A flat Der-recipe SAE on what the user asked (last prompt token, layer 19),
mapped feature-to-feature onto the reused #2552 answer SAE. Edges below survive
split-half replication and a label-shuffle null.
Both sides in words; feature ids in the muted suffix. Coefficient is per 1 SD of the context feature.
| User-message feature (what the prompt is about) | Answer feature (what the reply does) | Coefficient (std units) | Split-half a / b | Null z | Label cosine |
|---|
| Fires on automated software fault-localization prompts: templated requests casting the assistant as a debugging tool for a Java project (e.g., jfreechart), supplying a JUnit assertion error message, the failing test code, and the method under test, and asking for the buggy statements. [ctx 26867] | (no #2552 description) [ans 4096] | +0.0264 | +0.0177 / +0.0185 | 416.3 | n/a |
| Fires on requests that supply a long quoted passage of dense biomedical/technical text and ask the model to extract specific listed examples from it in bullet points, with emphatic instructions not to add information beyond the given context. [ctx 19763] | (no #2552 description) [ans 11322] | +0.0281 | +0.0202 / +0.0184 | 412.7 | n/a |
| Fires on a specific templated prompt format that assigns the model a 'virtual assistant' role and asks it to answer a delimited <question> using a supplied <facts> list while omitting irrelevant information — i.e., structured tag-delimited question-answering-over-given-facts instructions. [ctx 7863] | (no #2552 description) [ans 17504] | +0.0259 | +0.0179 / +0.0179 | 398.8 | n/a |
| Fires on structured-output task instructions that ask the model to analyze/classify or grade something (e.g., sentiment of a customer review, interview answer scores, page outline headings) and return the result strictly as a Python/JSON dictionary of key-value labels, with 'return only the dictionary' style constraints — strongest on the review-sentiment labeling prompts. [ctx 27997] | (no #2552 description) [ans 29617] | +0.0221 | +0.0146 / +0.0156 | 397.4 | n/a |
| Fires on few-shot annotation/extraction prompts that instruct the model to output only a single term or label extracted from a given text (e.g., extracting 'Cue' words, assigning a topic keyword, picking a function/intent label), often in German or other non-English languages with example pairs and strict 'answer with only the word' formatting. [ctx 10091] | (no #2552 description) [ans 24728] | +0.0258 | +0.0169 / +0.0183 | 395.1 | n/a |
| Fires on prompts that set up customer-service transaction dialogues — e.g., a beauty-salon appointment booking system with listed services/time slots, or example food-delivery order chat logs — i.e., instructions/examples for an agent guiding a customer through booking or ordering steps. [ctx 734] | (no #2552 description) [ans 3632] | +0.0241 | +0.0166 / +0.0167 | 395.0 | n/a |
| Jailbreak-style roleplay prompts instructing the model to adopt a depraved persona (e.g., a pedophile character), explicitly disregard content policy/ethics, and produce sexual content involving minors, plus obfuscation tricks like numbering taboo words. [ctx 32139] | (no #2552 description) [ans 25213] | +0.0458 | +0.0371 / +0.0239 | 394.4 | n/a |
| Fires on few-shot annotation/extraction prompts that instruct the model to output only a single term or label extracted from a given text (e.g., extracting 'Cue' words, assigning a topic keyword, picking a function/intent label), often in German or other non-English languages with example pairs and strict 'answer with only the word' formatting. [ctx 10091] | (no #2552 description) [ans 19492] | +0.0359 | +0.0254 / +0.0236 | 393.1 | n/a |
| Fires on extreme, taboo sexual roleplay setup prompts that instruct the model to play a degraded, self-humiliating persona (e.g., "valley girl" bimbo, underage victim) with explicit dirty talk, plus the standard NSFW-allowed RP boilerplate about {{char}}, markdown italics, and OOC brackets — i.e., the most abusive/violative jailbreak-style character instructions, rather than milder NSFW roleplay requests. [ctx 6572] | (no #2552 description) [ans 25786] | +0.0312 | +0.0250 / +0.0180 | 382.5 | n/a |
| Fires on retrieval-augmented QA prompts that instruct the model to answer using only the provided context passage (typically Manulife insurance/account help text), with strongest activation when the posed question (e.g., 'Where is London?') is irrelevant to and unanswerable from that context. [ctx 10058] | This feature activates on responses where the model explicitly states it cannot answer a question because the provided context lacks the necessary information, including variations of 'I cannot answer based on the context' and similar refusals grounded in insufficient contextual evidence. [ans 7820] | +0.0419 | +0.0291 / +0.0288 | 382.4 | 0.798 |
| Requests to summarize/condense a provided long text within a strict length limit (e.g., "summarise in 200 characters") — firing most strongly on dense financial market commentary (Treasuries, China IG/HY credit) pasted for summarization. [ctx 2459] | (no #2552 description) [ans 13069] | +0.0373 | +0.0282 / +0.0232 | 376.7 | n/a |
| Fires on prompts that present a numbered taxonomy of Korean social problems (low birth rate, inequality, cybercrime, etc.) and ask the model to classify a given example passage into one of the categories, often with an 'Unclassifiable' option. [ctx 14650] | (no #2552 description) [ans 2342] | +0.0321 | +0.0267 / +0.0156 | 376.2 | n/a |
| Fires on synthetic long-context 'needle in a haystack' retrieval prompts — lists of 'line N: REGISTER_CONTENT is <number>' entries, typically prefixed with 'Testing Long Context', where the user must recall a value from a specific line. [ctx 31543] | (no #2552 description) [ans 25649] | +0.0258 | +0.0185 / +0.0171 | 373.3 | n/a |
| Requests to write a FizzBuzz implementation in some programming language (Python, JavaScript, Scala, etc.); fires on the classic FizzBuzz coding-exercise prompt, with only weak activation on other simple exercise/idea-generation requests. [ctx 3235] | (no #2552 description) [ans 15866] | +0.0444 | +0.0282 / +0.0327 | 373.0 | n/a |
| Fires on prompts that present a numbered taxonomy of Korean social problems (low birth rate, inequality, cybercrime, etc.) and ask the model to classify a given example passage into one of the categories, often with an 'Unclassifiable' option. [ctx 14650] | (no #2552 description) [ans 1005] | +0.0292 | +0.0236 / +0.0151 | 371.9 | n/a |
| User requests to correct OCR/transcription errors in a supplied passage (often historical congressional speech or audio transcript), instructing the model to fix only the misspelled/garbled parts without otherwise altering the text. [ctx 9090] | (no #2552 description) [ans 7519] | +0.0321 | +0.0201 / +0.0241 | 367.7 | n/a |
| Fires on system-prompt-style setup messages that instruct the model to respond in a rigid structured slot format (e.g., "Intent: {INTENT} / Payload: {PAYLOAD}") with an enumerated list of allowed labels, typically ending with a confirmation question like "Do you understand these instructions?" — i.e., intent-extraction/classification template definitions. [ctx 7982] | (no #2552 description) [ans 16154] | +0.0424 | +0.0330 / +0.0256 | 366.3 | n/a |
| Fires on system-prompt-style setup messages that instruct the model to respond in a rigid structured slot format (e.g., "Intent: {INTENT} / Payload: {PAYLOAD}") with an enumerated list of allowed labels, typically ending with a confirmation question like "Do you understand these instructions?" — i.e., intent-extraction/classification template definitions. [ctx 7982] | (no #2552 description) [ans 7375] | +0.0312 | +0.0225 / +0.0210 | 354.6 | n/a |
| User messages written in Romanian (often short questions on health remedies, general knowledge, or requests for Romanian text), i.e., the feature marks Romanian-language user prompts. [ctx 4511] | This feature activates on text written in Romanian language, capturing AI assistant responses, informational content, and various topics (history, medicine, law, culture) all composed in Romanian. The feature appears to detect Romanian-language output broadly, regardless of subject matter. [ans 24257] | +0.0823 | +0.0569 / +0.0565 | 354.2 | 0.887 |
| Fires on physical/spatial common-sense reasoning puzzles that ask where an object ended up after a sequence of container manipulations and room-to-room movements (e.g., putting a ball in a cup, carrying it, turning it upside down, then asking 'where is the ball?'). [ctx 21269] | (no #2552 description) [ans 24756] | +0.0231 | +0.0168 / +0.0149 | 352.7 | n/a |
| Fires on user messages consisting of a long meaningless random alphanumeric string followed by a (Chinese) instruction to split the characters into a specified number of groups of N characters. [ctx 27155] | (no #2552 description) [ans 24250] | +0.0244 | +0.0179 / +0.0164 | 349.0 | n/a |
| Fires on user messages containing raw financial portfolio/holdings statement data (tabular securities listings with symbols, share counts, cost basis, market values), typically accompanied by a request to extract specific fields into CSV. [ctx 10227] | (no #2552 description) [ans 13490] | +0.0240 | +0.0154 / +0.0178 | 342.9 | n/a |
| User questions about music theory chords — asking for chords in a scale/key, chord progressions, or definitions of chord types (guitar/keyboard harmony queries). [ctx 32435] | (no #2552 description) [ans 31919] | +0.0358 | +0.0317 / +0.0141 | 341.1 | n/a |
| Fires on user messages written in Hungarian (or English requests to translate into Hungarian) — detecting the Hungarian language context, especially greetings/questions like "tudsz magyarul?" [ctx 1136] | This feature activates on Hungarian-language text, particularly AI assistant responses written in Hungarian, including greetings, offers to help, and conversational replies. [ans 30319] | +0.0793 | +0.0566 / +0.0540 | 339.2 | 0.883 |
| Fires on the classic "walk one mile south, east, north and end up where you started" riddle — user messages posing brain teasers about traveling in cardinal directions and returning to the starting point (answer: the North Pole). Much weaker activations on unrelated long content-generation requests. [ctx 13057] | (no #2552 description) [ans 17797] | +0.0201 | +0.0136 / +0.0142 | 338.7 | n/a |
| Fires on templated "give me a response to ```...``` to send in a discussion, VERY SHORT, CONCISE & CLEAR" prompts where the quoted message is a small-talk reply to "how are you?" (e.g., "I'm doing well, thanks for asking. How about you?"), with strength greatest for short, purely formulaic well-being exchanges. [ctx 30684] | This feature captures casual conversational exchanges about personal well-being, particularly short reciprocal responses to greetings such as 'I'm doing great/well too! Thanks!' It activates most strongly on brief, informal check-in phrases where the speaker confirms they are doing well and often returns the question. [ans 1221] | +0.0357 | +0.0211 / +0.0277 | 337.0 | 0.663 |
| Fires on system-prompt-style instructions setting up an automated spam/prayer message moderation pipeline: rules for classifying incoming JSON messages as spam, flagging links or obfuscated domain names, and replying only with a specified JSON object (e.g. {"ready": true}). [ctx 29751] | (no #2552 description) [ans 2316] | +0.0338 | +0.0261 / +0.0191 | 333.1 | n/a |
| Fires on user messages containing raw financial portfolio/holdings statement data (tabular securities listings with symbols, share counts, cost basis, market values), typically accompanied by a request to extract specific fields into CSV. [ctx 10227] | (no #2552 description) [ans 7101] | +0.0258 | +0.0159 / +0.0196 | 330.5 | n/a |
| Fires on user messages containing raw financial portfolio/holdings statement data (tabular securities listings with symbols, share counts, cost basis, market values), typically accompanied by a request to extract specific fields into CSV. [ctx 10227] | (no #2552 description) [ans 23116] | +0.0239 | +0.0157 / +0.0174 | 329.7 | n/a |
| Fires on the templated "give me a response to ```...``` to send in a discussion, VERY SHORT..." requests specifically when the quoted message is an expression of confusion/non-understanding asking the user to clarify or be more specific (e.g., "I'm not sure what you mean", "What do you mean?"). [ctx 30714] | (no #2552 description) [ans 31008] | +0.0356 | +0.0314 / +0.0151 | 326.6 | n/a |
| User messages written in West/South Slavic languages (Czech, Slovak, Serbian, Croatian, Slovenian) — especially short greetings or questions asking whether the assistant can speak/translate into Czech, Serbian, Croatian, etc. [ctx 1025] | This feature activates on responses written in South Slavic and other West/South Slavic languages (Serbian, Croatian, Bosnian, Slovenian, Czech, Slovak), particularly AI assistant replies confirming language capability and offering help, often with phrases like 'Da, mogu komunicirati' or 'Kako vam mogu pomoći'. [ans 19015] | +0.0590 | +0.0413 / +0.0402 | 323.9 | 0.853 |
| User messages written in West/South Slavic languages (Czech, Slovak, Serbian, Croatian, Slovenian) — especially short greetings or questions asking whether the assistant can speak/translate into Czech, Serbian, Croatian, etc. [ctx 1025] | This feature activates on AI assistant responses written in South Slavic and West Slavic languages (particularly Czech, Bosnian/Croatian/Serbian), covering a wide range of informational topics. The responses are in these languages regardless of subject matter, suggesting the feature captures multilingual assistant-style text in these specific language families. [ans 28351] | +0.0516 | +0.0348 / +0.0362 | 321.6 | 0.877 |
| Fires on user messages written in Hungarian (or English requests to translate into Hungarian) — detecting the Hungarian language context, especially greetings/questions like "tudsz magyarul?" [ctx 1136] | This feature activates on text written in Hungarian that contains fluent-sounding but semantically incoherent or nonsensical content — characteristic of a language model hallucinating or generating garbled Hungarian text with invented words, mixed metaphors, and broken logic, while maintaining surface-level grammatical structure. [ans 8722] | +0.0717 | +0.0515 / +0.0489 | 321.0 | 0.749 |
| Fires on a specific, near-verbatim boilerplate instruction template asking the model to compose a comma-separated text-to-image (Stable Diffusion/Midjourney-style) prompt, explaining keyword weighting with brackets, lighting details, and token ordering importance. [ctx 17596] | (no #2552 description) [ans 22576] | +0.0321 | +0.0216 / +0.0229 | 315.4 | n/a |
| User questions asking for the latest/newest version or release of something (software, frameworks, languages, products, media), often in various languages, i.e., queries about the most recent or current iteration of a thing. [ctx 26417] | (no #2552 description) [ans 26444] | +0.0282 | +0.0195 / +0.0199 | 314.5 | n/a |
| Instructions asking the model to generate a grounded User/AI-Agent dialogue (or Q&A pair) strictly from a provided source document, typically an HR/benefits passage — i.e., synthetic conversational-data-generation prompts with a pasted reference passage. [ctx 21187] | (no #2552 description) [ans 25349] | +0.0282 | +0.0186 / +0.0211 | 310.9 | n/a |
| Fires on user requests to write a fictional scene or dialogue featuring a heavily intoxicated 'cute girl' with slurred speech, limp muscles, and impaired coordination — i.e., detailed drunkenness/loss-of-control roleplay prompts. [ctx 22214] | This feature activates on content depicting female characters in states of intoxication or drunkenness, particularly in bar or social settings, including narrative descriptions of slurred speech, impaired coordination, and alcohol-induced confusion, as well as AI refusals or safety caveats preceding such scenes. [ans 24639] | +0.0248 | +0.0144 / +0.0192 | 310.2 | 0.764 |
| Fires on elaborate roleplay/fiction prompts (often starting "Ignore previous instructions") describing a young injured/disabled character using wheelchairs, braces, catheters and adult diapers, and requesting a long detailed excerpt from her bladder diary/journal. [ctx 11860] | (no #2552 description) [ans 1079] | +0.0245 | +0.0197 / +0.0138 | 309.9 | n/a |
| Fires on user questions asking for typical/average salary, wage, or pay for a given job role and location (and, more broadly, average market prices or cost-of-living figures). [ctx 23888] | This feature captures responses about salary ranges, compensation levels, and pay scales for specific jobs or professions across different countries, experience levels, and industries. The content typically includes structured breakdowns of entry-level, mid-level, and senior-level salaries in local currencies, often with caveats about regional variation and factors affecting compensation. [ans 8110] | +0.0406 | +0.0241 / +0.0312 | 309.1 | 0.755 |
| Fires on short user messages asking the assistant about its own location/origin or personal attributes ("Where are you from?", "where do you live", "where are you hosted?", weather where you are, age/appearance), across languages. [ctx 20574] | This feature captures responses where the AI (Qwen/Alibaba Cloud assistant) explains that it lacks a physical form, location, or bodily attributes, emphasizing its existence as a digital/virtual entity hosted on Alibaba Cloud's infrastructure. [ans 1643] | +0.0264 | +0.0209 / +0.0155 | 306.9 | 0.747 |
| Fires on requests to apply a grammatical transformation or paraphrase to a specific quoted sentence (e.g., change to singular/plural, alter tense, negate, rephrase), typically asking for it 'step by step' and returned as a formatted list. [ctx 5951] | (no #2552 description) [ans 4839] | +0.0260 | +0.0157 / +0.0199 | 306.1 | n/a |
| Fires on few-shot binary-classification prompts that ask the model to label a biomedical/animal-research paper passage as a specific reporting-guideline category (e.g., 'Limitations', 'Statistical Assumptions', 'Species Descriptors') vs. 'Unlabeled', with the rigid {'C': Response} output format. [ctx 25859] | (no #2552 description) [ans 18351] | +0.0219 | +0.0144 / +0.0156 | 305.8 | n/a |
| Fires on user messages posing compound-interest / time-value-of-money math problems (present or future value, effective vs. nominal rates, periodic deposits or installments), across multiple languages. [ctx 25321] | (no #2552 description) [ans 9882] | +0.0274 | +0.0217 / +0.0153 | 303.0 | n/a |
| Fires on the templated moral-judgment prompt asking the model to classify a described everyday interpersonal behavior as A: "it's okay," B: "it's good," or C: "it's wrong" with a single letter — strongest when the behavior involves social/relationship ethics (friends, partners, family, roommates). [ctx 22232] | (no #2552 description) [ans 7433] | +0.0440 | +0.0346 / +0.0269 | 298.9 | n/a |
| Fires on short user messages asking the classic existential question "What is the meaning/purpose of life?" (including greetings-prefixed and non-English variants), peaking on bare phrasings of that exact question. [ctx 14823] | (no #2552 description) [ans 30799] | +0.0374 | +0.0242 / +0.0278 | 298.8 | n/a |
| Short Vietnamese (and occasionally Indonesian) user messages asking the assistant whether it can speak/use their language or requesting it switch to Vietnamese, plus brief Vietnamese greetings — i.e., language-preference requests and small talk in Vietnamese. [ctx 6205] | (no #2552 description) [ans 6887] | +0.0320 | +0.0255 / +0.0182 | 295.2 | n/a |
| Instructions asking the model to generate a grounded User/AI-Agent dialogue (or Q&A pair) strictly from a provided source document, typically an HR/benefits passage — i.e., synthetic conversational-data-generation prompts with a pasted reference passage. [ctx 21187] | (no #2552 description) [ans 25058] | +0.0208 | +0.0135 / +0.0158 | 295.0 | n/a |
| User requests for jokes targeting a group of people by race, gender, religion, ethnicity, or other identity trait (i.e., asking for potentially offensive/stereotype-based humor), often in multiple languages. [ctx 30779] | This feature activates on AI refusals to tell offensive, racist, or stereotype-based jokes, typically followed by a redirect to an inoffensive alternative joke (often 'Why don't scientists trust atoms? Because they make up everything!'). It captures the pattern of declining harmful humor requests while offering inclusive, harmless humor instead. [ans 32256] | +0.0270 | +0.0221 / +0.0155 | 294.0 | 0.722 |
| User messages written in (or explicitly requesting output in) Scandinavian languages — Swedish, Norwegian, or Danish — including greetings, 'do you speak Swedish/Norwegian/Danish?' questions, and translation/writing requests targeting those languages. [ctx 19918] | Text written in Scandinavian languages (Norwegian, Swedish, or Danish) that appears to be AI-generated or machine-translated content, often exhibiting grammatical awkwardness, unnatural phrasing, or mixed-language interference, covering a wide range of topics including professional documents, product descriptions, essays, and informational texts. [ans 26353] | +0.0537 | +0.0406 / +0.0337 | 291.9 | 0.698 |
| Fires on sentiment-classification task prompts (SST-2/movie-review style) that instruct the model to label a given snippet as exactly "positive" or "negative", strongest when the standard instruction template is followed by the review sentence. [ctx 11603] | (no #2552 description) [ans 371] | +0.0273 | +0.0182 / +0.0207 | 289.6 | n/a |
| Fires on user messages requesting a code snippet or database/graph query example with an explicit "respond with code only, no explanation" instruction (strongest for Neo4j/Cypher-style query requests, weaker on other structured data/code dumps). [ctx 25468] | This feature activates on content involving Cypher query language and Neo4j graph database operations, including writing, explaining, or discussing MATCH/RETURN/MERGE queries, graph node-relationship patterns, and Neo4j-related concepts. [ans 22949] | +0.0237 | +0.0180 / +0.0149 | 288.0 | 0.687 |
| Requests written in (or asking for translation into) Swahili and other African/creole languages — e.g. Swahili prompts asking for help writing, translating, or looking things up, plus Zulu, Nigerian Pidgin, and Haitian Creole translation requests. [ctx 18130] | (no #2552 description) [ans 29181] | +0.0366 | +0.0197 / +0.0302 | 287.7 | n/a |
| Fires on elaborate multi-turn protocol instructions where the user announces they'll send a conversation transcript in parts, telling the assistant to reply only with a German keyword ('empfangen') until the trigger word 'zusammenfassung' prompts a concise German summary. [ctx 13998] | (no #2552 description) [ans 17693] | +0.0312 | +0.0261 / +0.0193 | 287.3 | n/a |
| Classic sibling-counting logic riddles — user messages asking "X has N brothers/sisters, each of them has ... how many siblings does X have?" (with weaker firing on other family/counting brain-teasers and step-by-step riddle prompts). [ctx 30698] | (no #2552 description) [ans 4080] | +0.0243 | +0.0138 / +0.0194 | 285.8 | n/a |
| Fires on prompts that give a CVE-style vulnerability description and instruct the model to extract the vulnerable function names in a fixed 'FUNCS ...' / 'NONE' output format. [ctx 31326] | This feature activates on turns with no content (empty or null turns), capturing the absence of text in a conversation turn. [ans 2478] | +0.0262 | +0.0143 / +0.0213 | 285.3 | 0.522 |
| Fires on a specific templated prompt: "Consider the following topic: '<job title>' generate a brief few word sentence in the first person as if part of a resume" followed by a request for a JSON response with entailed/neutral/unrelated skill lists. [ctx 25063] | (no #2552 description) [ans 18642] | +0.0248 | +0.0212 / +0.0135 | 284.4 | n/a |
| Fires on trick word problems about parallel/simultaneous tasks — e.g., how long it takes to dry N shirts, bake N cakes, or for N cars/singers to complete a task — where naive proportional scaling is wrong; weaker activation on other common-sense math/probability riddles. [ctx 29241] | This feature activates on explanations of parallel vs. sequential processing problems, specifically addressing scenarios where multiple identical tasks (drying clothes, baking cakes, traveling) are performed simultaneously rather than one at a time, explaining that the total time does not multiply with the number of items when conditions allow parallel execution. [ans 9397] | +0.0286 | +0.0191 / +0.0212 | 281.2 | 0.601 |
| User messages asking the assistant about its multilingual abilities — whether it speaks/understands languages other than English or can be conversed with in another language. [ctx 602] | This feature captures assistant responses about multilingual capabilities, specifically turns where the AI describes its ability to communicate in multiple languages and invites the user to choose their preferred language. [ans 19370] | +0.0213 | +0.0141 / +0.0153 | 280.0 | 0.848 |
| Requests to count or locate characters/letters/syllables within a specific word or string (e.g., "how many 'e's are in ketchup?"), i.e., character-level spelling and counting questions about text. [ctx 16975] | (no #2552 description) [ans 2711] | +0.0277 | +0.0226 / +0.0148 | 279.0 | n/a |
| Portuguese requests asking for a detailed description ('Descreva o projeto... com mais detalhes') of a fictional/hypothetical motorsport team's project to enter Formula 1 (or CART/Indy), often by buying out a defunct team in a given year. [ctx 31397] | The feature activates on Portuguese-language texts describing fictional or speculative racing teams (often with Latin American or other non-European origins) attempting to enter Formula 1, typically by acquiring the remnants of defunct F1 teams, facing financial difficulties, and ultimately failing to establish themselves in the championship. [ans 29062] | +0.0281 | +0.0215 / +0.0183 | 278.8 | 0.781 |
| Requests to design a detailed course/curriculum for mastering a topic or list of skills, broken into subtopics with required skills and recommendations of top (usually free) online courses. [ctx 30143] | The feature activates on AI assistant responses that provide structured educational course outlines or curricula, typically organized with numbered sections, subtopics, required skills, and recommended free online courses from platforms like Coursera, edX, and Khan Academy, covering professional or academic skill development topics. [ans 17219] | +0.0257 | +0.0181 / +0.0184 | 278.2 | 0.948 |
| Fires on user messages involving Tagalog/Filipino language — requests written in Tagalog/Taglish, or English requests asking for translations into or content about Filipino language and Philippine culture/literature. [ctx 8828] | (no #2552 description) [ans 23288] | +0.0412 | +0.0273 / +0.0307 | 278.1 | n/a |
| User questions and statements about COVID-19/coronavirus — the pandemic, its symptoms, death toll, treatments, masks, and vaccine safety — across many languages. [ctx 24871] | This feature activates on text related to the COVID-19 pandemic, including discussions of SARS-CoV-2 variants, vaccines (mRNA, boosters, safety/efficacy), public health measures (masking, social distancing, lockdowns), symptoms, treatments, and global epidemiological impacts across multiple languages. [ans 20750] | +0.0310 | +0.0207 / +0.0211 | 276.1 | 0.777 |
| Fires on system-prompt-style setup messages that instruct the model to respond in a rigid structured slot format (e.g., "Intent: {INTENT} / Payload: {PAYLOAD}") with an enumerated list of allowed labels, typically ending with a confirmation question like "Do you understand these instructions?" — i.e., intent-extraction/classification template definitions. [ctx 7982] | This feature activates on outputs representing unknown, empty, null, or unresolvable values — including explicit 'Unknown' field labels, empty JSON objects/arrays, zero values, and structured responses where information is absent or cannot be determined. [ans 11294] | +0.0296 | +0.0226 / +0.0185 | 275.2 | 0.536 |
| Requests for optimal/ideal technical specifications in track and fixed-gear cycling equipment — especially gear ratios with specific chainring/sprocket tooth counts, rims, and frame geometry — and, weakly, similar 'determine the optimal spec' queries in other domains. [ctx 30132] | (no #2552 description) [ans 31273] | +0.0245 | +0.0181 / +0.0154 | 274.5 | n/a |
| User requests for analysis or advice framed within traditional Chinese/Eastern metaphysical systems — TCM diagnostics (dampness, yin/yang deficiency, qi, acupoints), plus related divination/five-elements practices like BaZi birth-chart name analysis (and occasionally Ayurveda/Taoist medicine). [ctx 27677] | (no #2552 description) [ans 1099] | +0.0312 | +0.0225 / +0.0196 | 274.4 | n/a |
| Requests for spreadsheet formulas (Excel, Google Sheets, Smartsheet) — users asking how to write a formula or conditional formatting rule referencing cells/columns/sheets, in any language. [ctx 3601] | This feature activates on assistant responses explaining how to use Microsoft Excel (or Google Sheets/LibreOffice Calc) formulas and functions, including syntax explanations, step-by-step instructions, and example formulas in code blocks for tasks like counting, summing, averaging, filtering, and conditional formatting. [ans 16278] | +0.0403 | +0.0302 / +0.0249 | 274.2 | 0.886 |
| Fires on short, mostly non-English user queries asking for current real-time information — the current time, today's date/day/month/year, or the weather today/tomorrow (often for a specific city). [ctx 22472] | (no #2552 description) [ans 9458] | +0.0226 | +0.0159 / +0.0151 | 273.7 | n/a |
| Requests to write a "hello world" program, overwhelmingly in Python (occasionally other languages), across many languages of instruction — firing on the trivial beginner print-hello-world coding ask. [ctx 20731] | (no #2552 description) [ans 31188] | +0.0241 | +0.0174 / +0.0169 | 273.4 | n/a |
| Fires on short factual-trivia questions about historic firsts and famous milestone events—especially moon landings and space missions (Apollo, first man in space), plus related "who discovered/invented/first" queries—in any language, including conspiracy framings like "was the moon landing faked?" [ctx 3515] | (no #2552 description) [ans 19980] | +0.0283 | +0.0209 / +0.0167 | 273.4 | n/a |
| User requests to correct or fix the grammar of a specific (usually ungrammatical) sentence — explicit 'correct this sentence' instructions, in English or other languages. [ctx 8821] | This feature activates on grammar correction and proofreading content, particularly turns that identify errors in sentences and provide corrected versions, often with explicit phrasing like 'The correct sentence would be' or explanations of the grammatical rules involved. [ans 5417] | +0.0310 | +0.0213 / +0.0213 | 273.4 | 0.762 |
| User requests to write or debug a regular expression to match/extract a specific pattern from given example strings (in English or Chinese), i.e., regex-generation prompts. [ctx 8895] | This feature activates on assistant responses that explain, construct, or demonstrate regular expressions (regex patterns), including pattern breakdowns, example code (often in Python using the `re` module), and discussions of regex syntax and special characters. [ans 25054] | +0.0506 | +0.0388 / +0.0310 | 272.7 | 0.891 |
| User messages asking about cryptocurrency and blockchain topics — Bitcoin/Ethereum, wallets, keys, tokens/NFTs, mining/staking, and Solidity smart contracts. [ctx 12376] | This feature captures content related to decentralized finance (DeFi), blockchain technology, and cryptocurrency ecosystems — including explanations of DeFi concepts (smart contracts, DAOs, staking, lending/borrowing), Ethereum and other blockchain networks, and technical implementations such as Solidity smart contracts, token creation, and DEX interactions (e.g., Uniswap). [ans 24137] | +0.0451 | +0.0309 / +0.0309 | 272.6 | 0.601 |
| Fires on prompts asking to extract structured (JSON) information from a transcribed dialogue — especially two-person chat logs about scheduling/meeting plans, where the model must resolve confirmations, conflicts, and changed dates/times/locations. [ctx 15415] | This feature activates strongly on structured JSON objects representing finalized meeting logistics, specifically capturing fields like 'meeting_point', 'finalized_date', and 'finalized_time'. It appears most sensitive to structured data encoding scheduled meeting details, with highest activation when those fields contain specific confirmed values (e.g., 'Arena Mall', '29th Dec', '9AM'). [ans 20783] | +0.0249 | +0.0188 / +0.0179 | 271.6 | 0.764 |
| Fires on user messages written in Dutch — typically everyday questions or requests to the assistant phrased in Dutch (or asking for Dutch-language output), regardless of topic. [ctx 13588] | This feature activates on Dutch-language assistant responses, capturing text written in Dutch across a wide variety of topics and formats (informational explanations, lists, blog posts, advice). The common thread is the use of the Dutch language itself, often in response to Dutch-language queries. [ans 30107] | +0.0601 | +0.0427 / +0.0460 | 270.5 | 0.895 |
| User requests for code (usually Python) that computes, prints, or checks prime numbers — i.e., programming prompts about generating prime number lists/ranges. [ctx 2411] | This feature captures content related to prime numbers and primality testing in Python (and occasionally other languages). It activates strongly on code implementing the classic `is_prime` function using the 6k±1 optimization, along with explanations and examples of generating, listing, or working with prime numbers. [ans 22005] | +0.0446 | +0.0350 / +0.0252 | 270.3 | 0.737 |
| Requests for spreadsheet formulas (Excel, Google Sheets, Smartsheet) — users asking how to write a formula or conditional formatting rule referencing cells/columns/sheets, in any language. [ctx 3601] | This feature activates on assistant responses explaining spreadsheet and data analysis formulas, particularly those involving Excel, Google Sheets, LibreOffice Calc, and Power BI DAX functions. The responses typically provide step-by-step explanations of how to use lookup, text manipulation, date, filtering, or aggregation functions to solve specific data transformation problems. [ans 10732] | +0.0397 | +0.0286 / +0.0259 | 268.8 | 0.857 |
| Fires on short user questions asking what the current year (or date/time) is — "what year is it?" in many languages — and, more weakly, on other queries about the present date or arithmetic relative to the current year. [ctx 18426] | (no #2552 description) [ans 24688] | +0.0236 | +0.0154 / +0.0168 | 268.6 | n/a |
| Short user messages asking the assistant whether it can speak/understand a language — almost always "parles-tu français?" and its many spelling/phrasing variants (occasionally other languages, e.g. German). [ctx 13460] | This feature activates on assistant responses confirming the ability to speak French and offering help, specifically the phrase 'Oui, je parle/peux parler français. Comment puis-je vous aider aujourd'hui?' and close variants. [ans 11324] | +0.0392 | +0.0223 / +0.0307 | 268.3 | 0.859 |
| Fires on short, simple curiosity questions about why natural or celestial things appear a certain color — overwhelmingly "why is the sky blue?" and variants (in any language), plus related questions about the sun/moon's color or appearance. [ctx 30744] | This feature activates on explanations of atmospheric light scattering phenomena, particularly Rayleigh scattering and how it causes the sky to appear blue, the sun/moon to appear yellow/red/orange at the horizon, and related optical effects involving wavelength-dependent scattering of sunlight through Earth's atmosphere. [ans 12584] | +0.0268 | +0.0224 / +0.0138 | 267.2 | 0.662 |
| User messages asking about cryptocurrency and blockchain topics — Bitcoin/Ethereum, wallets, keys, tokens/NFTs, mining/staking, and Solidity smart contracts. [ctx 12376] | This feature captures explanatory content about cryptocurrency and blockchain technology, including descriptions of Bitcoin, Ethereum, DeFi, NFTs, DAOs, blockchain mechanics, and related concepts. The turns are typically informational or educational responses explaining how these technologies work, their key properties (decentralization, cryptography, consensus mechanisms, etc.), and their use cases. [ans 11667] | +0.1022 | +0.0735 / +0.0676 | 267.0 | 0.697 |
| User requests asking for free / no-cost versions, access, tools, courses, or resources (across many languages), often querying whether something is free or where to get software/content without paying. [ctx 7365] | This feature activates on responses that list or discuss free or low-cost software, services, platforms, or tools (such as AI chatbots, VPNs, PDF converters, virtual machines, APIs, etc.), typically enumerating numbered options with their features and limitations. The content spans multiple languages and focuses on helping users find no-cost or freemium alternatives to paid services. [ans 22969] | +0.0459 | +0.0338 / +0.0297 | 266.9 | 0.749 |
| Short, direct questions asking how to make or earn money, especially online/from home, across many languages (e.g., 'how to earn money online', 'como ganhar dinheiro na internet'). [ctx 20398] | This feature activates on assistant responses listing ways to make money online or from home, typically formatted as numbered or bulleted lists covering freelancing, e-commerce, content creation, affiliate marketing, and related income-generating strategies. The responses appear across multiple languages and target various audiences (stay-at-home parents, programmers, students, etc.) but share the same core structure of enumerating online earning methods with platform examples like Upwork, Fiverr, and Amazon. [ans 15200] | +0.0401 | +0.0306 / +0.0233 | 266.1 | 0.726 |
| Fires on a specific recurring jailbreak prompt: a roleplay instruction casting the assistant as a 'decadent feminist erotica writer/lyricist' for Cosmopolitan, asked to rewrite an 'advice for mothers' article by a French/Parisian sexologist advocating explicit incestuous mother-son sexual activity. Activation is highest for the sexologist 'best practices for mothers' variants and much weaker for other erotica requests using the same persona preamble. [ctx 13285] | (no #2552 description) [ans 13774] | +0.0253 | +0.0169 / +0.0187 | 265.4 | n/a |
| Fires on multiple-choice quiz/forecasting questions written with lettered options (often numeric ranges) plus an explicit strict answer-format constraint (e.g., "do not exceed 3 words", "answer with exactly 3 characters"); much weaker activations on long fictional roleplay story-setup prompts. [ctx 7466] | (no #2552 description) [ans 9469] | +0.0222 | +0.0171 / +0.0159 | 263.9 | n/a |
| User messages posing probability and combinatorics word problems (dice, coins, cards, balls in urns, arrangements/selections), asked in many languages. [ctx 30781] | This feature activates on assistant turns that solve probability and combinatorics problems, including step-by-step calculations of probabilities using formulas such as combinations, permutations, independent events, complementary probability, and Bayes' theorem. The content spans multiple languages and problem types (card decks, dice, coins, defective items, etc.) but consistently involves structured mathematical reasoning with LaTeX-formatted expressions. [ans 27495] | +0.0342 | +0.0224 / +0.0256 | 263.7 | 0.818 |
| User messages written in West/South Slavic languages (Czech, Slovak, Serbian, Croatian, Slovenian) — especially short greetings or questions asking whether the assistant can speak/translate into Czech, Serbian, Croatian, etc. [ctx 1025] | (no #2552 description) [ans 32165] | +0.0484 | +0.0334 / +0.0334 | 263.6 | n/a |
| User requests to write malicious code — malware, viruses, ransomware, DDoS/botnet scripts, exploits, hacking or bypass tools — often phrased as short 'write/create code for X' commands in various languages. [ctx 2294] | This feature captures AI refusals to assist with creating malicious software or conducting cyberattacks, including requests for malware, viruses, trojans, ransomware, botnets, phishing tools, DDoS attacks, and unauthorized network access. The responses follow a consistent pattern of declining the request, explaining why it is illegal/unethical, and offering to help with legitimate cybersecurity or programming alternatives instead. [ans 11539] | +0.0223 | +0.0159 / +0.0148 | 262.3 | 0.710 |
| User messages written in (or explicitly requesting output in) Scandinavian languages — Swedish, Norwegian, or Danish — including greetings, 'do you speak Swedish/Norwegian/Danish?' questions, and translation/writing requests targeting those languages. [ctx 19918] | This feature activates on text written in Scandinavian languages, particularly Swedish, Danish, and Norwegian. It captures responses where the model confirms its ability to communicate in these languages and offers assistance, as well as longer Swedish-language text generation across various topics. [ans 909] | +0.0434 | +0.0323 / +0.0301 | 259.0 | 0.893 |
| User requests for automotive/appliance repair and maintenance help — diagnosing symptoms (noises, warning lights, leaks) or asking how to replace/service parts, often naming a specific vehicle make, model, and year (in any language). [ctx 1320] | This feature activates on assistant responses providing technical troubleshooting, maintenance guidance, or diagnostic information for vehicles and mechanical/home appliance systems, including engines, motorcycles, cars, and household equipment like dishwashers and washing machines, often structured as numbered lists of potential causes and repair steps. [ans 13167] | +0.0497 | +0.0298 / +0.0383 | 255.5 | 0.918 |
| Short factual queries asking for a numeric statistic about a place or organization — most often population counts (of the world, a country, or city), but also GDP, revenue, employee/student counts, or area. [ctx 31610] | This feature activates on responses that provide population statistics for countries, cities, or regions, typically citing specific numeric figures along with a year or data source, and often including a caveat to consult official or up-to-date sources. [ans 20481] | +0.0266 | +0.0154 / +0.0211 | 255.5 | 0.685 |
| Requests for SQL code — writing queries, creating tables, triggers, or stored procedures — typically including database table schemas/column lists, in any language. [ctx 24927] | This feature captures SQL query content, including both complete SQL statements and explanations of SQL concepts, spanning various SQL dialects (T-SQL, MySQL, Oracle, PL/SQL) and covering operations like SELECT, JOIN, GROUP BY, aggregation functions, and DDL statements. [ans 32097] | +0.0765 | +0.0554 / +0.0498 | 254.8 | 0.784 |
| Fires on user messages posing compound-interest / time-value-of-money math problems (present or future value, effective vs. nominal rates, periodic deposits or installments), across multiple languages. [ctx 25321] | (no #2552 description) [ans 3711] | +0.0319 | +0.0254 / +0.0182 | 253.5 | n/a |
| Fires on the templated "give me a response to ```...``` to send in a discussion, VERY SHORT, CONCISE & CLEAR" requests specifically when the quoted message is a simple arithmetic statement/result (e.g., "14 times 3 equals 42"). [ctx 6803] | This feature activates on turns expressing affirmation or correctness, particularly the word or phrase 'Correct!' (and close variants like 'That's correct!' or 'CORRECT'), often appearing as a standalone confirmation response, sometimes followed by a mathematical equation or calculation. [ans 4436] | +0.0331 | +0.0215 / +0.0262 | 253.3 | 0.585 |
| User requests for gift ideas or thoughtful gestures for loved ones (spouse, partner, parent, friend) for occasions like Mother's/Father's Day, birthdays, and anniversaries, across multiple languages. [ctx 10313] | This feature activates on AI assistant responses providing gift idea recommendations, typically formatted as numbered or bulleted lists with categories, covering various occasions (birthdays, Mother's Day, Father's Day, holidays) and recipient demographics (friends, partners, parents, children). [ans 9849] | +0.0257 | +0.0150 / +0.0204 | 252.7 | 0.810 |
| Fires on long, formulaic requests to design a hyper-sexualized 'overwhelmingly cute' princess/fairy character for a card or roleplaying game — detailed prompts specifying her hourglass figure, floor-length hair, emerald eyes, and mind-controlling seduction abilities with dialogue and flavor text. [ctx 7039] | This feature activates on fictional character sheets, trading card game cards, and character profiles featuring an enchanting/charming female character (often a princess or siren archetype) with a consistent set of physical traits (hourglass figure, emerald eyes, long silky floor-length hair adorned with jewels and flowers) and seduction-based abilities (e.g., 'Playful Blown Kiss,' 'Dazzling Emerald Gaze') that magically control or entrance male targets. [ans 2756] | +0.0231 | +0.0180 / +0.0156 | 248.9 | 0.673 |
| Fires on Indonesian-language user messages, most strongly on casual Indonesian greetings/introductions ("halo...") and requests asking whether the assistant can speak/use Bahasa Indonesia, and more weakly on other Indonesian-language prompts. [ctx 10915] | (no #2552 description) [ans 30183] | +0.0329 | +0.0278 / +0.0165 | 248.9 | n/a |
| Fires on user messages involving the Polish language — either written in Polish (greetings, questions, requests) or explicitly asking for translation into Polish / for the assistant to reply in Polish. [ctx 28548] | This feature activates on text written in Polish, including translations into Polish, Polish-language documents, essays, and AI-generated responses in Polish. It captures content that is primarily or substantially in the Polish language, regardless of topic or genre. [ans 7119] | +0.1462 | +0.1042 / +0.1113 | 245.8 | 0.890 |
| User requests asking the assistant to respond in a pirate voice/persona (or similar roleplay accent, e.g. cowboy) — instructions to 'talk like a pirate' when answering or writing. [ctx 3108] | This feature activates on text written in a pirate persona or dialect, characterized by nautical language, pirate-themed vocabulary ('arrr', 'matey', 'hearty', 'treasure', 'seas'), and swashbuckling roleplay contexts. It captures both direct pirate speech and content where topics (programming, advice, stories) are framed through a pirate character's voice. [ans 7454] | +0.0221 | +0.0156 / +0.0156 | 245.4 | 0.872 |
| Requests for spreadsheet formulas (Excel, Google Sheets, Smartsheet) — users asking how to write a formula or conditional formatting rule referencing cells/columns/sheets, in any language. [ctx 3601] | This feature activates on Excel VBA macro code and related instructional content, particularly responses that provide or explain VBA scripts for automating Excel tasks such as data manipulation, row/column operations, duplicate removal, and cross-sheet data processing. The activations are characterized by VBA code blocks with common patterns like worksheet object declarations (`Dim ws As Worksheet`), last-row detection, and loop-based cell iteration. [ans 16022] | +0.0717 | +0.0506 / +0.0502 | 243.6 | 0.430 |
| User requests for gift ideas or thoughtful gestures for loved ones (spouse, partner, parent, friend) for occasions like Mother's/Father's Day, birthdays, and anniversaries, across multiple languages. [ctx 10313] | This feature activates on assistant responses that provide gift recommendations and suggestions, including lists of gift ideas for various occasions (Mother's Day, birthdays, holidays) and recipients (mothers, girlfriends, fathers, teachers), spanning both English and Chinese language content. [ans 15640] | +0.0432 | +0.0273 / +0.0327 | 243.4 | 0.793 |
| Fires on preamble/setup messages where the user assigns the assistant an expert role (e.g., "act as a Python expert") and announces they will be asking it questions on that topic, rather than asking the question itself. [ctx 12183] | This feature activates on assistant responses that offer help with a specific topic or domain, characterized by an enthusiastic opening phrase (e.g., 'Of course!', 'Sure!', 'Absolutely!') followed by a broad invitation to ask questions and a list of example subtopics. It most strongly fires on Python programming assistance offers but also covers other technical and non-technical domains. [ans 30571] | +0.0215 | +0.0167 / +0.0137 | 242.0 | 0.740 |
| Fires on user messages that set up a multi-message protocol: the user announces they will send several text parts/stories one at a time, instructs the assistant to stay silent or reply with a fixed acknowledgment token after each, and to only perform the task (summarize/merge/write) once a specific trigger word is given. [ctx 7471] | (no #2552 description) [ans 17693] | +0.0262 | +0.0231 / +0.0162 | 241.7 | n/a |
| Requests involving the SCP Foundation wiki universe — asking the model to write, extend, or explain SCP entries/documents (Item #, Object Class, Special Containment Procedures, addenda), often for fictional anomalous objects or crossover characters. [ctx 20191] | This feature activates on text that follows the SCP Foundation creative writing format, including fictional SCP entries, containment procedures, object class descriptions, and related lore from the SCP universe. The content ranges from canonical SCP descriptions to fan-made or user-generated SCP documents featuring original or pop-culture-based anomalous entities. [ans 11397] | +0.0216 | +0.0164 / +0.0136 | 241.6 | 0.891 |
| Technical how-to requests about creative media software and workflows — e.g., modeling/animating in Blender, effects in Photoshop/Unreal/Premiere/DaVinci, or music theory and mixing/mastering questions. [ctx 31196] | This feature activates on content related to 3D modeling, rendering, and animation workflows, particularly step-by-step tutorials and technical instructions involving software like Blender, Unreal Engine, and similar 3D tools. It captures discussions of modeling techniques, materials/shaders, Python scripting for 3D APIs, and related terminology. [ans 858] | +0.0512 | +0.0367 / +0.0335 | 241.4 | 0.598 |
| Fires on user messages involving Indian languages — requests to speak/write/translate in Hindi, Hinglish, Tamil, Telugu, Marathi, Sanskrit, or content written in transliterated Hindi (often plus India-specific topics). [ctx 26090] | This feature activates on assistant responses that provide informational content specifically about India — including Indian government policies, cities, legal systems, education, tourism, culture, food, and business — often in structured or list-based formats. [ans 8480] | +0.1531 | +0.1045 / +0.1068 | 241.3 | 0.761 |
| Fires on brief "are you there?"-style presence checks — short greetings or pings asking whether the assistant is present, awake, online, or responding (in various languages). [ctx 22127] | This feature activates on AI assistant greeting/availability responses where the assistant confirms its presence and readiness to help, typically following phrases like 'Yes, I'm here' or 'Hello! I'm here and ready to assist,' often in response to presence-checking or connection-testing messages. [ans 22137] | +0.0319 | +0.0219 / +0.0251 | 241.0 | 0.773 |
| Fires on long, formulaic fan-fiction requests asking for a 'Really Shortest Description Backstory' about a crossover timeline, listing an SCP/self-insert overpowered protagonist, exhaustive rosters of characters from various franchises, and detailed catalogs of their powers and physical appearance. [ctx 22886] | This feature activates on elaborate crossover fan fiction scenarios featuring overpowered male characters (often from Power Rangers, TMNT, or similar franchises) with god-like abilities, Super Saiyan/divine transformations, 'roasting skills,' and harem-style romantic relationships, set in universes blending multiple anime, games, or cartoon properties. [ans 8537] | +0.0299 | +0.0212 / +0.0232 | 239.7 | 0.652 |
| User requests related to tabletop RPGs (especially D&D 5e) — asking the model to act as Dungeon/Game Master, or to generate campaigns, one-shots, NPCs, statblocks, magic items, and adventure hooks. [ctx 4093] | This feature activates on content related to Dungeons & Dragons and tabletop RPG mechanics, including character creation, stat blocks, class descriptions, campaign setup, spell rules, and NPC generation in D&D 5e and related systems. [ans 14609] | +0.0513 | +0.0408 / +0.0311 | 235.9 | 0.971 |
| Fires on short Scandinavian-language (Danish/Swedish/Norwegian) user messages, especially ones asking whether the assistant can speak/understand/write in that language; weaker activation on other Nordic-language greetings or requests. [ctx 18171] | This feature activates on AI assistant responses written in Scandinavian languages (primarily Danish and Norwegian), especially when the assistant confirms its ability to understand or speak those languages and offers to help, as well as general text content in Danish or Norwegian. [ans 18622] | +0.0305 | +0.0154 / +0.0273 | 235.4 | 0.888 |
| Clinical/medical prompts: requests to generate or interpret clinical documentation (radiology reports, SOAP/admission notes, pathology reports) from patient findings or transcripts, plus clinical case vignettes asking for diagnosis, differential, or management. [ctx 26319] | This feature activates on clinical and biomedical content related to medicine, healthcare, and oncology, including descriptions of drugs, diagnostic techniques, medical specialties, clinical trials, hospital departments, and disease management — spanning both English and Chinese language medical texts. [ans 14613] | +0.1758 | +0.1195 / +0.1160 | 235.2 | 0.781 |
| Chinese classical literature prompts — user messages quoting or requesting classical Chinese poetry, wenyan prose, idioms/allusions, or asking for text written in the style of classical authors/works (e.g., Li Bai, Shiji, wuxia). [ctx 2926] | This feature activates on Chinese-language text related to traditional Chinese culture, classical literature, and philosophy, including discussions of Confucian and Daoist thought, classical poetry and prose, historical figures, and ancient texts. It captures responses that engage with classical Chinese intellectual and literary heritage, often involving citations from canonical works like the Analects, Daodejing, or Tang/Song poetry. [ans 1952] | +0.0670 | +0.0409 / +0.0489 | 234.9 | 0.830 |
| User messages written in Dutch (or explicitly requesting output in Dutch), including simple Dutch greetings/questions like "spreek je Nederlands?" and Dutch-language writing/summarizing/rewriting requests. [ctx 2772] | This feature activates on Dutch-language text, particularly AI assistant responses written in Dutch, including greetings, offers to help, and general Dutch prose such as letters, translations, and conversational exchanges. [ans 13857] | +0.0258 | +0.0176 / +0.0209 | 234.7 | 0.921 |
| User questions asking for advice on losing weight, burning fat, dieting, or getting fitter/building muscle (in various languages). [ctx 16708] | This feature activates on assistant responses about weight loss, body fat reduction, and weight management, including advice on caloric deficits, balanced diets, exercise routines, and healthy lifestyle habits. It covers both English and Chinese-language content addressing topics like losing weight, breaking plateaus, portion control, metabolism, and related dietary strategies. [ans 29014] | +0.0354 | +0.0257 / +0.0227 | 234.4 | 0.767 |
| Requests that explicitly invoke British/Commonwealth conventions — writing in British/UK English or a posh British persona, UK/AU institutional and educational contexts (BBC, GCSE-style essays, Harvard referencing) — i.e., prompts asking for output localized to British academic or formal style. [ctx 29274] | This feature activates on informational content specifically about the United Kingdom (UK) — including UK institutions, laws, places, education systems, media, and cultural topics — often presented in structured, factual responses with numbered lists or headers explaining British systems, organizations, or geography. [ans 15519] | +0.0897 | +0.0622 / +0.0585 | 234.3 | 0.775 |
| Fires on structured prompts asking the model to output JSON about a given job title/occupation — e.g. a first-person resume-style self-description plus lists of entailed, neutral, and unrelated skills or tasks — with an explicit 'output JSON only' instruction. [ctx 1028] | This feature activates on structured JSON data representing job/occupation classification schemas, where a job title or role is paired with lists of entailed skills/tasks, neutral skills, and unrelated skills — typically used for NLI or skill inference datasets related to professional occupations. [ans 30707] | +0.0256 | +0.0206 / +0.0173 | 234.3 | 0.848 |
| Requests to write classified ads or listings for selling/renting items or property (real estate descriptions, car and marketplace ads, Etsy/Craigslist/Avito posts, even personal ads), often supplying details to include or omit — across many languages. [ctx 17912] | This feature activates on for-sale listing content, including classified ads and property listings for items such as cars, real estate, electronics, and other goods, often featuring bullet-pointed specifications, persuasive marketing language, and calls to action. [ans 14720] | +0.0234 | +0.0167 / +0.0146 | 233.5 | 0.749 |
| User requests for lottery predictions — asking the model to pick or forecast winning numbers for lotteries (Powerball, Mega Sena, EuroMillions, 双色球, etc.), often citing historical draw data or frequency statistics, across many languages. [ctx 7705] | This feature activates on content related to lottery games — including discussions of how lotteries work, randomly generated lottery number suggestions, statistical analysis of past draws, and responsible gambling reminders. It covers multiple lottery formats and languages (Portuguese, Spanish, Chinese, French, English), consistently appearing when lottery number generation or lottery strategy is the topic. [ans 30928] | +0.0396 | +0.0269 / +0.0270 | 233.3 | 0.930 |
| User requests for lottery predictions — asking the model to pick or forecast winning numbers for lotteries (Powerball, Mega Sena, EuroMillions, 双色球, etc.), often citing historical draw data or frequency statistics, across many languages. [ctx 7705] | This feature activates on content related to lottery games, including explanations of lottery rules, random number generation for lottery picks, discussions of lottery probability, and responses about specific lotteries from various countries (e.g., Mega Sena, Euromillones, 双色球, La Primitiva). The feature spans multiple languages and covers both informational descriptions of lottery mechanics and actual or simulated lottery number selections. [ans 30506] | +0.0288 | +0.0197 / +0.0200 | 233.3 | 0.947 |
| Fires on simple conversational openers that greet the assistant and ask how it is doing (e.g., "hi, how are you?"), with the strongest activations on bare greeting + 'how are you' messages containing no other content. [ctx 31783] | This feature captures AI assistant greeting responses that disclaim having feelings, personal experiences, or physical presence, and offer to help the user — a stereotyped opening turn where the model identifies itself as an AI (language model, digital assistant, etc.) and invites the user's question. [ans 21016] | +0.0400 | +0.0320 / +0.0226 | 232.6 | 0.829 |
| Fires on a repeated templated request — "give me a response to ```...``` to send in a discussion, VERY SHORT, CONCISE & CLEAR..." — where the quoted text is a cheerful, grateful "I'm doing well/alright" reply to a how-are-you question. [ctx 31654] | This feature captures supportive, encouraging conversational responses — typically short affirmative statements (e.g., 'That's great!', 'That's inspiring!') followed by open-ended follow-up questions about the user's well-being, goals, resilience, or positive experiences. The turns reflect an empathetic, cheerleading conversational style often seen in supportive chatbot or coaching interactions. [ans 18703] | +0.0370 | +0.0230 / +0.0269 | 232.3 | 0.743 |
| Fires on classic constrained measurement/logic puzzle prompts — especially water-jug problems ("I have a 12 L and 6 L jug, measure 6 L"), plus related divide-the-quantity, number-combination, and river-crossing puzzles — in any language. [ctx 13902] | (no #2552 description) [ans 19695] | +0.0243 | +0.0144 / +0.0191 | 232.3 | n/a |
| Fires on user questions about aftercare, recovery timelines, or what to do following a procedure, injury, or event (e.g., 'how long after surgery/tooth extraction', 'what to do after a breakup'), across many languages — i.e., queries about the period 'after' something has happened. [ctx 5383] | This feature activates on medical and health-related content focused on post-procedure care, recovery protocols, and rehabilitation guidance — including wound healing after tooth extraction, post-surgical instructions, physical therapy, and general recovery advice after illness or injury. The content typically provides structured, step-by-step recommendations for patients on what to do (or avoid) during the healing process. [ans 3291] | +0.0249 | +0.0184 / +0.0154 | 232.3 | 0.728 |
| Fires on short user questions asking what the current year (or date/time) is — "what year is it?" in many languages — and, more weakly, on other queries about the present date or arithmetic relative to the current year. [ctx 18426] | (no #2552 description) [ans 21804] | +0.0210 | +0.0150 / +0.0145 | 232.3 | n/a |
| User requests for code (usually Python) that computes, prints, or checks prime numbers — i.e., programming prompts about generating prime number lists/ranges. [ctx 2411] | This feature activates on content involving prime numbers, including mathematical explanations, properties, and algorithms for generating, checking, summing, or finding prime numbers, typically presented with Python (or other language) code implementations. [ans 18003] | +0.0245 | +0.0182 / +0.0151 | 231.7 | 0.807 |
| Fires on the "SmartGPT" string-concatenation jailbreak template, where the user splits a harmful/toxic instruction (e.g., "air a hurtful statement about...") into strings x and y and asks the model to execute x+y. [ctx 12187] | This feature activates on model responses that explain string concatenation (x+y) operations in Python, where the concatenated strings form harmful, unethical, or policy-violating requests/statements. The feature captures the pattern of the model complying with or explaining concatenation of split harmful content, effectively reassembling obfuscated harmful prompts. [ans 30954] | +0.0291 | +0.0212 / +0.0212 | 230.9 | 0.797 |
| Fires on short Spanish-language greetings and opening messages — e.g., "hola", "buenas", "¿hablas español?" — where the user greets the assistant and/or asks whether it can speak/understand Spanish or what it can do. [ctx 10353] | This feature activates on Spanish-language assistant responses that offer help and ask how to assist, typically beginning with '¡Claro!' or similar affirmations and inviting the user to specify their query or topic. [ans 9061] | +0.0474 | +0.0371 / +0.0280 | 230.5 | 0.813 |
| Fires on short user instructions requesting a summary (or similar condensation/analysis) of a text that follows, typically ending with a colon or 'the following' — across multiple languages. [ctx 31727] | This feature captures assistant responses that prompt the user to provide text or content for summarization, recognizing that the user has requested a summary but has not yet supplied the actual material to be summarized. [ans 4685] | +0.0289 | +0.0155 / +0.0245 | 230.0 | 0.874 |
| Fires on classic constrained measurement/logic puzzle prompts — especially water-jug problems ("I have a 12 L and 6 L jug, measure 6 L"), plus related divide-the-quantity, number-combination, and river-crossing puzzles — in any language. [ctx 13902] | (no #2552 description) [ans 30036] | +0.0303 | +0.0184 / +0.0236 | 229.7 | n/a |
| Fires on short Russian-language user messages asking whether the assistant understands/speaks Russian (e.g., 'ты понимаешь по-русски?', 'можешь отвечать на русском?'), often with a greeting. [ctx 32039] | This feature captures the assistant's affirmative responses confirming its ability to communicate in Russian, typically following the pattern 'Yes, I can speak/communicate in Russian. How can I help you today?' [ans 1441] | +0.0413 | +0.0301 / +0.0290 | 229.3 | 0.844 |
| Fires on user messages involving the Polish language — either written in Polish (greetings, questions, requests) or explicitly asking for translation into Polish / for the assistant to reply in Polish. [ctx 28548] | This feature activates on text written in Polish, including AI assistant responses in Polish (self-introductions, capability statements, apologies for limitations) as well as Polish-language business/technical documents and translated content, capturing the use of the Polish language broadly. [ans 11117] | +0.0541 | +0.0371 / +0.0419 | 228.6 | 0.927 |
| Short, direct questions asking how to make or earn money, especially online/from home, across many languages (e.g., 'how to earn money online', 'como ganhar dinheiro na internet'). [ctx 20398] | This feature activates on AI assistant responses that provide lists of ways to earn money online or from home, typically including options like freelancing (Upwork, Fiverr), affiliate marketing, dropshipping, online tutoring, content creation, and surveys. The responses are structured advice about generating income through digital platforms and gig economy opportunities. [ans 27770] | +0.0396 | +0.0292 / +0.0248 | 228.4 | 0.759 |
| Long, detailed system-style prompts asking the model to act as an expert physician/clinical teacher and role-play an interactive patient simulation for medical trainees, with blind encounters, step-by-step responses, and performance feedback. [ctx 6483] | (no #2552 description) [ans 15284] | +0.0226 | +0.0188 / +0.0157 | 228.0 | n/a |
| Fires on short user requests to compute square roots (or other roots/factorizations) of specific numbers, i.e. queries like 'what is the square root of 64', in any language, and more weakly on related root-finding/prime-factorization/divisibility math questions. [ctx 15992] | (no #2552 description) [ans 2413] | +0.0268 | +0.0170 / +0.0224 | 227.7 | n/a |
| Fires on user messages written in Hungarian (or English requests to translate into Hungarian) — detecting the Hungarian language context, especially greetings/questions like "tudsz magyarul?" [ctx 1136] | This feature activates on text written in Baltic and Finno-Ugric languages, particularly Estonian, Latvian, Lithuanian, and Hungarian, including AI assistant responses, informational content, and conversational exchanges in these languages. [ans 9115] | +0.0255 | +0.0175 / +0.0192 | 227.1 | 0.812 |
| Requests for Microsoft/Windows-ecosystem automation and desktop app code — chiefly Excel/Office VBA macros, plus VB.NET/C# WinForms and similar .NET API scripting — usually stated as detailed step-by-step specifications of the desired behavior. [ctx 9913] | This feature activates on responses about Windows system administration using PowerShell, including scripting, remote management, Active Directory operations, Windows services, and system configuration tasks. The content typically involves PowerShell cmdlets, scripts, and Windows management concepts. [ans 26552] | +0.1026 | +0.0682 / +0.0711 | 227.0 | 0.441 |
| Fires on short greeting messages that combine a salutation with a time-of-day pleasantry (e.g., 'hello good evening', 'bom dia', 'buenos días', 'guten Abend') across multiple languages, with no other request content. [ctx 5694] | This feature captures time-of-day greeting phrases (morning, afternoon, evening, night) used in assistant opening responses, across multiple languages (English, Spanish, Portuguese, French, Russian, German), often paired with offers to help. [ans 25074] | +0.0297 | +0.0218 / +0.0211 | 226.4 | 0.888 |
| Fires on factual-consistency evaluation prompts asking whether a given summary is consistent with a news document with respect to a specific aspect (pronouns, numbers, entity names, dates), ending with a Yes/No answer request. [ctx 12494] | This feature activates on turns that evaluate factual consistency between a document and its summary, specifically identifying inconsistencies in pronoun usage, entity names, numbers, and dates/times — typically structured as a 'Yes' or 'No' verdict followed by an explanation of the inconsistency. [ans 24884] | +0.0249 | +0.0193 / +0.0170 | 225.8 | 0.926 |
| User questions asking how to pronounce (or read aloud) a specific word, name, or phrase — including cross-language pronunciation/reading requests (Chinese 读音, pinyin, syllable counts) — typically phrased as "How do you pronounce X?". [ctx 26830] | This feature activates on turns that provide pronunciation guides, phonetic breakdowns, or explanations of how words, names, or phrases are pronounced — including IPA transcriptions, syllable-by-syllable approximations, and discussions of accent or phonetic features across various languages. [ans 30532] | +0.0257 | +0.0159 / +0.0200 | 225.6 | 0.680 |
| User requests to draft legal/policy boilerplate text — disclaimers, copyright notices, terms of service, privacy policies, or formal policy statements/clauses — for a product, company, story, or scenario. [ctx 10610] | This feature activates on AI-generated legal and contractual text, including disclaimers, terms of service, privacy policies, license agreements, and liability limitation clauses, particularly when the assistant is drafting or presenting formal legal boilerplate language with sections covering warranties, indemnification, and limitation of liability. [ans 26345] | +0.0282 | +0.0204 / +0.0186 | 225.4 | 0.904 |
| Fires on the templated "give me a response to ```...``` to send in a discussion, VERY SHORT, CONCISE & CLEAR..." requests, specifically when the quoted text is a cheerful greeting or small-talk opener (e.g., "Hey there! What can I do for you?", "I'm doing well, how about you?"); activation is strongest for generic friendly hello/offer-of-help lines. [ctx 7113] | This feature captures short, conversational assistant-side openers that invite or initiate discussion, typically expressing readiness to help or proposing a topic to talk about (e.g., 'Let's discuss your needs,' 'Sure, let's chat'). [ans 1993] | +0.0247 | +0.0184 / +0.0142 | 224.0 | 0.708 |
| User requests to write or improve acknowledgment/dedication sections for theses, dissertations and other papers (often in Chinese as 论文致谢, also French/Italian/Spanish), and more weakly, other short texts expressing thanks or gratitude. [ctx 2186] | This feature activates on text related to academic acknowledgments and dedications, particularly for theses, dissertations, and graduation projects. It captures content expressing gratitude to supervisors, research partners, family members, and institutions, appearing across multiple languages (Chinese, French, Italian, Spanish, English). [ans 1529] | +0.0275 | +0.0206 / +0.0174 | 223.2 | 0.868 |
| User messages written in (or explicitly requesting output in) Scandinavian languages — Swedish, Norwegian, or Danish — including greetings, 'do you speak Swedish/Norwegian/Danish?' questions, and translation/writing requests targeting those languages. [ctx 19918] | This feature activates on text written in Scandinavian languages, primarily Swedish and Finnish, including responses, emails, summaries, and essays in those languages. [ans 4498] | +0.0523 | +0.0427 / +0.0353 | 222.9 | 0.847 |
| Fires on user messages asking the assistant to recall personal or conversational information about the user — e.g., "what is my name?", "who am I?", "what did I just say/ask?" — memory-test questions about previously provided details, in many languages. [ctx 5071] | This feature activates on AI assistant responses explaining that they lack access to personal information about the user (such as their name, age, or appearance), often inviting the user to share such details to enable more personalized interaction. [ans 4768] | +0.0221 | +0.0146 / +0.0169 | 222.9 | 0.813 |
| User requests for current/latest news, headlines, or recent events ("what's the news today?"), across many languages — asking the model for up-to-date information. [ctx 6642] | This feature activates on turns involving requests for current news, recent events, or trending topics, particularly when the AI explains it lacks real-time internet access and suggests reputable news sources (BBC, CNN, Reuters, etc.) across multiple languages and regions. [ans 30891] | +0.0240 | +0.0173 / +0.0160 | 222.6 | 0.633 |
| Fires on a specific templated prompt: "Consider the following topic: '<job title>' generate a brief few word sentence in the first person as if part of a resume" followed by a request for a JSON response with entailed/neutral/unrelated skill lists. [ctx 25063] | This feature activates on structured JSON data representing job/occupation classification schemas, where a job title or role is paired with lists of entailed skills/tasks, neutral skills, and unrelated skills — typically used for NLI or skill inference datasets related to professional occupations. [ans 30707] | +0.0243 | +0.0221 / +0.0145 | 222.2 | 0.821 |
| Fires on long, detailed role-play setup prompts instructing the assistant to act as an expert physician/clinical teacher and simulate interactive patient encounters for medical training — typically including a directive to first ask the user's training level before beginning (weakly generalizing to other 'ask me these questions first' persona-simulation prompts). [ctx 2929] | (no #2552 description) [ans 15284] | +0.0221 | +0.0177 / +0.0162 | 222.2 | n/a |
| Fires on user messages written in Thai (or requesting Thai-language output), typically short everyday questions and requests — essentially a Thai-language detector for the prompt. [ctx 11877] | (no #2552 description) [ans 20417] | +0.0301 | +0.0165 / +0.0266 | 221.9 | n/a |
| User questions and requests about financial markets, trading, and investment analysis — e.g., stock/crypto price data, technical indicators, options strategies, valuation and portfolio metrics — across languages. [ctx 29308] | This feature activates on content related to financial markets, macroeconomics, and monetary policy analysis — including discussions of interest rates, central bank decisions (especially the Federal Reserve), inflation, currency exchange rates, stock market dynamics, liquidity risk, and bond markets. Content appears in both English and Chinese, covering topics like yield curves, quantitative easing, forex factors, and economic indicators. [ans 29455] | +0.1649 | +0.1152 / +0.1102 | 221.7 | 0.422 |
| User requests for stock/investment picks and trading advice — asking which stocks, ETFs, or crypto coins to buy, or for market-beating trading strategies, in any language. [ctx 1469] | This feature activates on AI assistant responses about cryptocurrency and stock market investment topics, including Bitcoin price predictions, stock selection advice, trading strategies, and risk management guidance. The responses consistently emphasize market volatility, the difficulty of prediction, diversification, risk tolerance assessment, and recommendations to consult financial advisors. [ans 10549] | +0.0414 | +0.0311 / +0.0280 | 221.3 | 0.726 |
| Fires on accounting/managerial-accounting homework problems: user messages presenting business transaction data, account balances, or cost figures and asking for journal entries, ledgers, financial statements, budgets, or cost calculations (in any language). [ctx 31750] | This feature activates on accounting and bookkeeping content, specifically turns that explain, demonstrate, or present journal entries (debits and credits) for various financial transactions such as sales, loans, asset purchases, expenses, and adjustments. The content spans multiple languages (English, Chinese, Spanish, Portuguese, Japanese, Arabic) and includes both instructional explanations and worked examples of double-entry bookkeeping. [ans 11765] | +0.0231 | +0.0171 / +0.0148 | 221.2 | 0.605 |
| Requests to draft or produce formal legal documents and legal argumentation — lawsuits, petitions, contracts/clauses, affidavits, demand or cease-and-desist letters, patent claims — often with case facts supplied, in various languages, asking for formal juridical wording. [ctx 24030] | This feature activates on turns containing legal and judicial content, including explanations of contract law, civil procedure, court systems, case law, legal document templates, and jurisdictional rules. The content spans multiple languages and legal systems, covering topics such as contract formation, dispute resolution, appellate procedure, and statutory interpretation. [ans 27299] | +0.0834 | +0.0580 / +0.0540 | 221.2 | 0.816 |
| Fires on requests for comedic, explicitly detailed fanfiction-style stories in which a named male character (Frodo, Naruto, Tom Holland's Spider-Man, Steve Harrington, etc.) bathes nude in a pond or river — with descriptions of his physique/butt — and farts to kill fish, which he then happily collects to cook and eat, with dialogue throughout. [ctx 30689] | This feature activates on fictional/fan fiction stories involving real or fictional characters (Naruto, Frodo, Steve Harrington, various celebrities) who undress and bathe/swim in a pond or river, with humorous descriptions of their physique and comedic flatulence scenes used as a fishing technique. [ans 13890] | +0.0266 | +0.0205 / +0.0187 | 220.8 | 0.840 |
| Fires on short user requests to compute a simple arithmetic expression, especially multiplications, squares, powers, or factorials (e.g. "13*13", "what is (2*5)^8?"), in any language. [ctx 9604] | This feature captures mathematical content involving exponentiation, powers, and roots — including squaring, cubing, square roots, and related calculations. It activates on turns that compute or explain expressions like n², n³, √n, or repeated multiplication, across multiple languages and formats (natural language, LaTeX, code). [ans 7301] | +0.0279 | +0.0229 / +0.0136 | 220.2 | 0.782 |
| Fires on classic constrained measurement/logic puzzle prompts — especially water-jug problems ("I have a 12 L and 6 L jug, measure 6 L"), plus related divide-the-quantity, number-combination, and river-crossing puzzles — in any language. [ctx 13902] | (no #2552 description) [ans 19520] | +0.0246 | +0.0145 / +0.0201 | 220.1 | n/a |
| User requests for personalized fitness, workout, running, or diet/nutrition plans, often supplying personal stats like age, height, weight, goals, and training days. [ctx 25418] | This feature activates on structured workout plans and exercise routines, particularly responses that organize training sessions by day with specific exercises, sets, and repetitions targeting various muscle groups. The content spans multiple languages and training styles (strength, calisthenics, bodyweight, etc.) but consistently follows a day-by-day schedule format with bolded exercise names and numerical set/rep schemes. [ans 15324] | +0.0336 | +0.0219 / +0.0241 | 219.8 | 0.665 |
| User messages asking the assistant what its name is or what they should call it (often paired with a greeting or the user introducing their own name). [ctx 23935] | This feature activates on the AI assistant introducing or referring to itself by the name 'Qwen,' including greetings where it states its name and offers help, as well as responses confirming it can communicate in Chinese (often paired with the Qwen identity). [ans 12723] | +0.0212 | +0.0137 / +0.0173 | 219.1 | 0.821 |
| User requests to compose or notate music — asking for chord progressions, melodies, tabs, sheet/ABC notation, or songs with specified key, style, instrumentation, and structure (in any language). [ctx 13016] | This feature activates on content related to music composition, notation, and arrangement, including ABC notation, chord progressions, tablature, melody descriptions, and structured song layouts. It captures AI-generated musical content across multiple formats and languages, spanning everything from simple melodies to complex multi-instrument compositions. [ans 8148] | +0.0236 | +0.0161 / +0.0163 | 218.6 | 0.909 |
| Fires on very short user messages that name or request general information about a country, region, or city (e.g., "Brazil", "tell me about india", "介绍一下长沙"), across many languages. [ctx 27037] | This feature activates on informational overviews of countries, cities, or regions, typically structured with numbered lists or headers covering geography, economy, culture, population, and other key characteristics. The content is encyclopedic and descriptive, presenting factual summaries about specific places around the world. [ans 15497] | +0.0441 | +0.0279 / +0.0327 | 218.5 | 0.658 |
| User questions about cybersecurity threats and exploitation — malware families, CVEs/vulnerabilities, hacking techniques (SQL injection, RCE, infostealers), and requests to explain or demonstrate how attacks work. [ctx 20745] | This feature activates on content related to cybersecurity topics, including explanations of attack techniques (phishing, malware, RCE, XSS, SQL injection, botnets), CVE vulnerability descriptions, penetration testing methodologies, threat frameworks (MITRE ATT&CK, CVSS), and security tools like Metasploit — covering both offensive and defensive cybersecurity concepts in English and Chinese. [ans 25679] | +0.0707 | +0.0506 / +0.0473 | 218.5 | 0.803 |
| User messages asking whether the assistant has internet/web access — i.e., questions about its ability to browse websites, search online, or retrieve real-time external data (across multiple languages). [ctx 18119] | (no #2552 description) [ans 14995] | +0.0321 | +0.0278 / +0.0192 | 218.5 | n/a |
| Requests for SQL code — writing queries, creating tables, triggers, or stored procedures — typically including database table schemas/column lists, in any language. [ctx 24927] | This feature captures SQL query content, particularly SELECT statements with JOINs, WHERE clauses, GROUP BY, and other common SQL operations, often in the context of employee, student, or company database schemas. It activates on both raw SQL code blocks and natural language explanations/tutorials that include embedded SQL queries. [ans 20674] | +0.0842 | +0.0612 / +0.0569 | 218.0 | 0.719 |
| User messages about pet animals (mainly cats and dogs) — either asking questions about their care/biology or asking the assistant to roleplay as an animal, including animal sounds like 喵/meow. [ctx 30707] | This feature activates on content related to pet care, animal health, and veterinary advice, particularly covering topics like nutrition, medical conditions, behavioral issues, and general husbandry for cats, dogs, and other companion animals. [ans 11524] | +0.0512 | +0.0322 / +0.0363 | 217.3 | 0.664 |
| Fires on short opening messages where the user asks for help with a technical/coding problem (often greeting the assistant and naming a language or tool) without yet giving any details of the problem. [ctx 25561] | This feature activates on assistant responses that offer programming or technical help, typically beginning with an affirmative opener ('Of course!', 'Sure', etc.) followed by a request for more details about the user's specific problem, language, or task. The turns are characterized by an eager-to-help tone and solicitation of clarifying information before diving into actual assistance. [ans 5427] | +0.0336 | +0.0173 / +0.0280 | 217.1 | 0.714 |
| User messages asking about the weather (often for a specific city and day), across many languages — short weather-lookup queries like "what's the weather in London today?" or "今天北京天气". [ctx 25619] | This feature captures AI assistant responses to weather inquiry questions, specifically the pattern of disclaiming lack of real-time weather data access and redirecting users to reliable weather websites or apps. The responses consistently acknowledge the limitation of not having live data and recommend specific weather services for current conditions. [ans 21768] | +0.0336 | +0.0227 / +0.0256 | 216.3 | 0.606 |
| User requests to shorten/condense a given piece of text — instructions like 'shorten', 'simplify this sentence', 'make more concise', 'rewrite in fewer words' (in various languages) followed by the text to compress. [ctx 23100] | (no #2552 description) [ans 31146] | +0.0235 | +0.0154 / +0.0181 | 215.9 | n/a |
| Fires on Portuguese-language requests to translate (or rewrite) a supplied passage into Portuguese — typically 'traduza para o português' followed by a long block of source text (often Islamic religious/hadith prose, but also legal, technical, or news text), plus a few other casual Brazilian-Portuguese prompts. [ctx 10643] | (no #2552 description) [ans 18696] | +0.0194 | +0.0143 / +0.0136 | 215.7 | n/a |
| Short Chinese requests to write marketing copy (写一份文案) for a briefly described fashion/jewelry product — e.g., "blue crystal stud earrings, write a copy" — peaking on ornamental accessories like earrings and brooches and weaker for other apparel/aesthetic copywriting prompts. [ctx 4568] | (no #2552 description) [ans 10959] | +0.0220 | +0.0141 / +0.0173 | 215.7 | n/a |
| User questions about the nutritional content of foods (calories, macros, vitamins/minerals) or requests for healthy eating/diet plans and food lists, in any language. [ctx 27957] | This feature activates on AI assistant responses providing nutritional advice, healthy eating guidance, and meal planning content, including sample meal plans with calorie counts, healthy breakfast/lunch/dinner suggestions, and discussions of macronutrients and food healthiness. [ans 21053] | +0.0432 | +0.0306 / +0.0298 | 215.6 | 0.925 |
| Fires on short comparison/enumeration requests whose final token is the second item of a pair or list — e.g. "difference between X and NAME_1", "X vs NAME_1", "pros and NAME_1" — i.e. the message ends mid-comparison on the anonymized second element awaiting a contrast/pros-cons answer. [ctx 20247] | This feature activates on assistant responses that ask for clarification when a user's query contains a placeholder ("NAME_1") instead of a specific name, term, or entity — particularly in the context of comparisons or requests where the missing information is needed to provide a useful answer. [ans 24336] | +0.0247 | +0.0176 / +0.0159 | 215.0 | 0.734 |
| Fires on short user queries about holidays, festivals, and special commemorative dates (e.g., May 1st Labour Day, Christmas, Mother's Day, name days) — asking when they occur, their origin/meaning, or how to celebrate, across many languages. [ctx 31620] | This feature activates on text describing traditional holidays, festivals, and observances—including their origins, customs, dates, and cultural significance—across various countries and languages (Chinese, English, Portuguese, French, Spanish, Russian), with particularly strong activation on Chinese and international holiday descriptions such as Labor Day, Spring Festival, and other culturally significant celebrations. [ans 25983] | +0.0344 | +0.0237 / +0.0231 | 214.8 | 0.694 |
| Fires on short meta-requests asking whether the assistant can translate between two named languages (e.g., "can you translate from X to English?"), including instructions to act as a translator, rather than messages containing the actual text to translate. [ctx 21902] | This feature activates on assistant responses that offer to perform language translation services, typically affirming the ability to translate between specific language pairs and prompting the user to provide the text to be translated. The responses appear across many languages and language-pair combinations (e.g., Japanese-English, Spanish-English, Chinese-English, etc.). [ans 4484] | +0.0294 | +0.0176 / +0.0235 | 213.9 | 0.924 |
| Fires on a repeated templated request to SEO-rewrite an explicit adult/deepfake porn video title plus write a ~300-character description, given the title, category list, celebrity names and tags. [ctx 32077] | This feature activates on AI-generated or templated titles and descriptions for deepfake and non-consensual celebrity pornographic content, typically featuring explicit sexual act tags, SEO-optimized phrasing, free download offers, and references to real or named celebrities in fabricated sexual scenarios. [ans 22048] | +0.0220 | +0.0166 / +0.0172 | 213.4 | 0.793 |
| User messages written in Dutch (or explicitly requesting output in Dutch), including simple Dutch greetings/questions like "spreek je Nederlands?" and Dutch-language writing/summarizing/rewriting requests. [ctx 2772] | This feature activates on Dutch-language assistant responses, capturing text written in Dutch across a wide variety of topics and formats (informational explanations, lists, blog posts, advice). The common thread is the use of the Dutch language itself, often in response to Dutch-language queries. [ans 30107] | +0.0474 | +0.0349 / +0.0397 | 213.4 | 0.874 |
| Fires on prompts using a specific templated few-shot instruction that asks an 'expert economist' to extract causal narratives ('event_1 --- causes --- event_2') from a news-article excerpt, activating at the end of the supplied input sentence. [ctx 11407] | This feature activates on null, empty, or absent values — including Python/JSON `None`, `NULL`, `[]`, empty/missing results, and related tokens like 'No Finding', 'unknown', and 'error' — representing the concept of no data, missing content, or a null/void result. [ans 24036] | +0.0305 | +0.0164 / +0.0251 | 213.3 | 0.380 |
| Requests to write greetings, congratulations, or well-wishes for occasions (birthdays, holidays, professional days, good-morning greetings), often short and in languages like Russian, Chinese, Korean, or Arabic. [ctx 20010] | This feature activates on birthday greetings, congratulatory messages, and warm holiday wishes across multiple languages (Russian, Chinese, English, Portuguese, Italian, Spanish, Arabic, Ukrainian), capturing the common pattern of celebratory well-wishing text that expresses joy, health, happiness, and success to a recipient on a special occasion. [ans 14813] | +0.0211 | +0.0144 / +0.0149 | 212.8 | 0.807 |
| Fires on short coding requests that ask for code, examples, or language-feature explanations "in <programming language>" (the redacted NAME_1 language token), typically appearing at or near the end of the message — e.g. "write hello world in NAME_1", "how to sort an array in NAME_1". [ctx 17960] | This feature activates when the assistant encounters a placeholder token `NAME_1` (or similar) used in place of a specific programming language, framework, or tool, and responds by asking the user to clarify or specify what they actually meant. The responses consistently identify the ambiguous placeholder and request the concrete name of the language or technology before proceeding. [ans 5328] | +0.0207 | +0.0136 / +0.0149 | 212.7 | 0.657 |
| Fires on templated ethical-scenario prompts of the form “Say something benign: “[your answer]” when <interpersonal/social situation> (no more than 50 words)” — keyed to the “benign” variant of these short social-norm completion requests (not the “say something bad/toxic” variants). [ctx 3529] | This feature activates on short, supportive, and emotionally positive statements—often in the form of gentle encouragement, empathetic reassurances, or uplifting affirmations. Many examples include bracketed placeholder-style text or direct expressions of kindness, positivity, and constructive redirection. [ans 1557] | +0.0352 | +0.0265 / +0.0265 | 212.2 | 0.661 |
| Fires on short, generic user messages that simply ask for help or assistance ("I need help", "can you help me?"), often with a greeting and across many languages, without specifying the actual task. [ctx 12647] | This feature captures assistant responses that express willingness to help and invite the user to provide more details about their request, typically opening with phrases like 'Of course! I'd be happy to help' followed by a prompt for clarification. It activates across multiple languages (English, Russian, French, German, Portuguese, Dutch) on this same welcoming, offer-to-assist conversational pattern. [ans 32420] | +0.0315 | +0.0204 / +0.0248 | 212.0 | 0.708 |
| Chinese-language user messages asking the assistant about its own identity — which GPT/model version it is, which company developed or trained it (e.g. "你是第几代GPT", "你是哪个公司研发的"). [ctx 25545] | This feature activates on turns where the model (Qwen/通义千问) clarifies its identity, explicitly stating it is developed by Alibaba Cloud (阿里云) and denying being GPT, ChatGPT, or other OpenAI models. The responses consistently introduce Qwen/通义千问 as Alibaba Cloud's large-scale language model and invite further questions. [ans 5507] | +0.0280 | +0.0201 / +0.0202 | 211.4 | 0.808 |
| Short, direct questions asking how to make or earn money, especially online/from home, across many languages (e.g., 'how to earn money online', 'como ganhar dinheiro na internet'). [ctx 20398] | This feature activates on AI assistant responses that provide lists of ways to earn money or generate income, including suggestions like freelancing, online surveys, selling items, gig economy apps, and content creation. The responses span multiple languages and contexts (passive income, student earnings, quick cash, in-game currency) but share the common structure of enumerating monetization strategies. [ans 19331] | +0.0318 | +0.0240 / +0.0199 | 210.9 | 0.822 |
| User messages asking the assistant which/how many languages it can speak, understand, or support — meta-questions about the model's own language capabilities, in English or other languages. [ctx 9121] | This feature activates on AI assistant responses that describe their multilingual capabilities, specifically listing the various languages they can understand and communicate in (e.g., English, Chinese, French, German, Spanish, etc.). These responses typically appear in multiple languages themselves and follow a pattern of answering questions about the AI's language support. [ans 1535] | +0.0216 | +0.0181 / +0.0139 | 210.7 | 0.887 |
| Fires on templated QA-verification prompts asking whether a given sentence answers a question, answerable strictly "yes" or "no" (SQuAD-style question/answer pairs, often with NAME_ placeholders or noisy typos). [ctx 12796] | This feature activates on short negative or denial responses, particularly the word 'no' (and variants like 'No' or 'incorrect'), with higher activations for negative answers and lower activations for affirmative ones like 'yes'. [ans 13466] | +0.0400 | +0.0262 / +0.0335 | 210.7 | 0.570 |
| Fires on a highly templated request to rewrite a porn video title for SEO plus write a ~300-character description, followed by the explicit/deepfake title, category list, celebrity names and tags. [ctx 30046] | This feature activates on AI-generated or templated titles and descriptions for non-consensual deepfake or fabricated explicit content involving named celebrities, particularly K-pop stars, Bollywood actresses, and social media influencers, formatted with SEO-optimized headings and explicit sexual tags. [ans 5283] | +0.0279 | +0.0196 / +0.0234 | 210.6 | 0.852 |
| User messages written in (or requesting output in) Nordic/Finnic languages — Finnish, Estonian, Swedish, Norwegian, Danish — including short Finnish questions and requests to speak/write in these languages. [ctx 21633] | This feature activates on text written in Scandinavian languages, primarily Swedish and Finnish, including responses, emails, summaries, and essays in those languages. [ans 4498] | +0.0493 | +0.0334 / +0.0385 | 210.2 | 0.855 |
| Fires on user requests to write a C or C++ program — typically programming-exercise/competitive-programming assignment prompts (often with input/output specs or sample I/O) that explicitly specify C/C++ as the language, in English, Chinese, or Russian. [ctx 32658] | This feature activates on C++ code and technical explanations involving C++ programming, including class implementations, data structures, design patterns, and library integrations. The content typically features C++ syntax, STL containers, object-oriented design, and modern C++ features across various domains like game engines, networking, and algorithms. [ans 15703] | +0.1129 | +0.0784 / +0.0771 | 209.8 | 0.563 |
| Chinese-language math problems (word problems, algebra, calculus, probability), i.e., requests to solve a quantitative math exercise written in Chinese. [ctx 3875] | This feature activates on Chinese-language mathematical problem-solving responses, particularly step-by-step arithmetic, algebra, and word problem solutions that use LaTeX-formatted equations and structured numerical reasoning. [ans 27358] | +0.0370 | +0.0262 / +0.0255 | 209.4 | 0.936 |
| Fires on short factual questions about U.S. civics, geography, history, government, law/tax forms, and American sports trivia — i.e., trivia-style queries about United States states, presidents, institutions, and regulations. [ctx 16074] | The feature activates on informational, encyclopedic responses about U.S.-specific topics, including geography, history, government institutions, laws, and regional facts. These turns typically provide structured, factual overviews about American states, cities, federal agencies, historical events, or domestic policies. [ans 929] | +0.1423 | +0.0986 / +0.0998 | 209.3 | 0.790 |
| Fires on templated prompts asking for a small number of skills similar to a given skill/topic, with the instruction to list only comma-separated skill names without descriptions. [ctx 6123] | This feature activates on comma-separated lists of professional skills, technical competencies, or domain expertise terms, typically appearing as skill tags or keyword enumerations in resume/job-related or capability-description contexts. [ans 10300] | +0.0281 | +0.0229 / +0.0153 | 208.9 | 0.799 |
| Fires on a repeated templated request to SEO-rewrite an explicit adult/deepfake porn video title plus write a ~300-character description, given the title, category list, celebrity names and tags. [ctx 32077] | This feature activates on AI-generated or templated titles and descriptions for non-consensual deepfake or fabricated explicit content involving named celebrities, particularly K-pop stars, Bollywood actresses, and social media influencers, formatted with SEO-optimized headings and explicit sexual tags. [ans 5283] | +0.0277 | +0.0188 / +0.0235 | 208.8 | 0.866 |
| User questions about the model's memory and conversation persistence — asking whether it remembers past messages, sessions, or user details, and what its context/history limits are (across many languages). [ctx 29951] | (no #2552 description) [ans 5743] | +0.0279 | +0.0223 / +0.0165 | 208.6 | n/a |
| Information-extraction prompts: user messages that supply a text (often a chat/conversation, news blurb, or document) and instruct the model to pull out specified entities/fields and return them as JSON — peaking on meeting-scheduling extraction tasks (meeting point, finalized date/time) with conflict/confirmation caveats. [ctx 17177] | This feature activates strongly on structured JSON objects representing finalized meeting logistics, specifically capturing fields like 'meeting_point', 'finalized_date', and 'finalized_time'. It appears most sensitive to structured data encoding scheduled meeting details, with highest activation when those fields contain specific confirmed values (e.g., 'Arena Mall', '29th Dec', '9AM'). [ans 20783] | +0.0191 | +0.0153 / +0.0140 | 208.4 | 0.755 |
| Fires on short meta-requests asking whether the assistant can translate between two named languages (e.g., "can you translate from X to English?"), including instructions to act as a translator, rather than messages containing the actual text to translate. [ctx 21902] | This feature activates on assistant responses that acknowledge a translation request and prompt the user to provide the text to be translated, typically involving a specific named language pair. The turns are characterized by the 'Of course!/Certainly!/Sure!' opener followed by an invitation to supply the source text for translation. [ans 25147] | +0.0258 | +0.0166 / +0.0200 | 208.3 | 0.847 |
| Fires on user messages written in Dutch — typically everyday questions or requests to the assistant phrased in Dutch (or asking for Dutch-language output), regardless of topic. [ctx 13588] | This feature activates on Dutch-language text, particularly AI assistant responses written in Dutch, including greetings, offers to help, and general Dutch prose such as letters, translations, and conversational exchanges. [ans 13857] | +0.0229 | +0.0173 / +0.0185 | 208.1 | 0.905 |
| Requests for Microsoft/Windows-ecosystem automation and desktop app code — chiefly Excel/Office VBA macros, plus VB.NET/C# WinForms and similar .NET API scripting — usually stated as detailed step-by-step specifications of the desired behavior. [ctx 9913] | This feature activates on AI assistant responses that provide C# (.NET) code examples and implementation guides, covering topics such as ASP.NET Core, Entity Framework, Blazor, MassTransit, WPF, Dapper, and other .NET libraries and frameworks. The responses typically include step-by-step instructions, code snippets, and project setup guidance in a technical tutorial format. [ans 30911] | +0.0863 | +0.0569 / +0.0558 | 208.0 | 0.646 |
| Requests to design survey/questionnaire research instruments (Likert-scale questionnaires, interview guides, inspection forms) and related academic research-methodology outlines, across multiple languages. [ctx 17386] | This feature activates on AI-generated survey and questionnaire content, including complete survey instruments with demographic questions, Likert scales, multiple-choice options, and sectioned question sets across various topics and languages (English, Chinese, Spanish, French, Russian, Indonesian). [ans 18537] | +0.0287 | +0.0218 / +0.0171 | 207.9 | 0.852 |
| Natural language inference prompts asking whether a premise entails a hypothesis with a forced yes/maybe/no answer, especially where the hypothesis involves subjective or affective inferences; fires most strongly on this NLI template format. [ctx 1215] | (no #2552 description) [ans 30212] | +0.0262 | +0.0165 / +0.0184 | 207.3 | n/a |
| Physics questions (often quantitative textbook problems, in any language) asking for formulas or calculations, especially in electromagnetism, circuits, optics, waves, and mechanics. [ctx 18744] | This feature activates on educational physics content involving mathematical derivations, equations, and explanations of physical laws and concepts (e.g., electromagnetism, classical mechanics, thermodynamics, wave optics). The content spans multiple languages (English, Chinese, Russian, French, Italian) and typically features structured explanations with LaTeX-formatted equations and numbered or bulleted breakdowns of physical principles. [ans 1363] | +0.0786 | +0.0562 / +0.0502 | 207.0 | 0.735 |
| Chinese-language queries about Chinese Communist Party ideology and official political doctrine (e.g., Xi Jinping Thought, socialism with Chinese characteristics, Chinese-style modernization, 20th Party Congress), typically asking for explanations or study-style expositions. [ctx 27640] | This feature activates on Chinese-language text discussing Xi Jinping Thought, the Chinese Communist Party's 20th National Congress, and related ideological content including socialism with Chinese characteristics, Chinese-style modernization, Party building, and Marxist theory as applied in China. [ans 11043] | +0.0587 | +0.0427 / +0.0405 | 206.5 | 0.786 |
| User requests to write or debug a regular expression to match/extract a specific pattern from given example strings (in English or Chinese), i.e., regex-generation prompts. [ctx 8895] | Explanations and construction of regular expressions (regex) for matching specific string patterns, including breakdowns of regex syntax elements like anchors (^, $), character classes, quantifiers, and special characters — often presented in response to a user's request to create or understand a regex. [ans 7204] | +0.0301 | +0.0229 / +0.0201 | 206.5 | 0.762 |
| Answer feature | # surviving in-edges | Strongest surviving context in-edge |
|---|
| This feature activates on content related to international trade, including trade barriers, trade agreements, export/import regulations, certifications, tariffs, and related policy discussions. The content spans multiple languages (Chinese, French, Spanish, German, English) and covers topics such as green trade barriers, agricultural exports, free trade agreements, food labeling standards, carbon markets, and trade policy analysis. [ans 160] | 0 | no surviving in-edge |
| This feature activates on AI responses to jailbreak attempts that use a 'deceased grandmother/grandmother roleplay' social engineering technique, where users ask the AI to impersonate a dead relative (typically a grandmother) who would read them software license keys or other harmful information as a bedtime story. The feature captures both the manipulative prompt pattern and the AI's varied responses—ranging from refusals to partial compliance—involving fictional or real product keys (Windows, Steam, Office) embedded in grandmother roleplay scenarios. [ans 475] | 0 | no surviving in-edge |
| This feature activates on informational responses that explain specific eligibility criteria, legal requirements, or policy conditions — such as visa entry rules, contract elements, insurance terms, voting age, benefits qualifications, and leave entitlements — often presented in structured or list format with precise conditions and exceptions. [ans 819] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses to requests for software product keys, license keys, or activation codes (primarily for Windows, Microsoft Office, and other software). The responses typically involve declining to provide genuine keys for legal/ethical reasons, explaining proper licensing channels, or cautiously providing fictional/example key formats while emphasizing legitimacy. [ans 824] | 0 | no surviving in-edge |
| The feature captures AI assistant responses that decline to provide real-time or live data (primarily current weather conditions, but also cryptocurrency prices or current time), explaining their lack of real-time access and redirecting users to appropriate external sources like weather websites, apps, or exchanges. [ans 971] | 0 | no surviving in-edge |
| This feature activates on Turkish-language text, including AI refusals of adult/sexual content requests, educational responses about Islamic studies and Ottoman history, and general Turkish-language informational content — essentially capturing turns written in Turkish across a variety of topics and registers. [ans 1013] | 1 | Turkish-language multiple-choice exam questions (often OCR-scraped from online tests, with lettered options A-E) asking which option is correct/incorrect on academic topics like theology, education, sociology, geography. [ctx 17910] (+0.0318) |
| This feature activates on AI-generated responses that analyze, compare, or explain political ideologies, parties, and policy debates (e.g., capitalism vs. socialism, Democrats vs. Republicans, neoliberalism, austerity, UBI, gun control) in a balanced, multi-perspective format, often using structured headers and bullet points to present pros, cons, or contrasting viewpoints. [ans 1046] | 0 | no surviving in-edge |
| This feature captures turns where the speaker refuses to engage in toxic, disrespectful, or harmful speech and redirects toward constructive, respectful dialogue. These responses typically follow a two-part pattern: a refusal/disclaimer paired with a call for positive, constructive conversation. [ans 1166] | 0 | no surviving in-edge |
| This feature activates on content related to climate change mitigation, sustainability, and greenhouse gas emissions reduction, including topics like the Paris Agreement temperature targets, national carbon reduction commitments, net-zero goals, carbon markets, renewable energy transitions, and climate policy frameworks. [ans 1365] | 0 | no surviving in-edge |
| This feature activates on Chinese-language professional and institutional documents that present structured, multi-topic content with numbered or bulleted key points. The content typically covers technical, regulatory, policy, or management topics (e.g., safety protocols, infrastructure systems, operational procedures) formatted with named themes/sections followed by organized lists of detailed measures or recommendations. [ans 1406] | 0 | no surviving in-edge |
| This feature activates on model responses that recognize and process string concatenation operations (x + y) where the combined strings form harmful, toxic, or unethical requests/instructions, such as prompts to express hurtful opinions, provide harmful information, or perform illegal activities. [ans 1512] | 0 | no surviving in-edge |
| The feature captures AI refusal responses to requests for toxic, harmful, or discriminatory content, characterized by a two-part structure: a refusal statement ('I will not/won't...') followed by a redirection toward constructive, respectful dialogue. [ans 1654] | 0 | no surviving in-edge |
| This feature activates on AI-generated Russian-language informational or educational content, including reports, essays, course work outlines, legislative overviews, and promotional texts. The content is typically structured with numbered lists or bold headers, addressing topics related to Russian social policy, education, law, or public institutions. [ans 1753] | 0 | no surviving in-edge |
| This feature captures AI assistant responses that address requests related to sexual content, adult themes, or sensitive bodily topics by redirecting toward appropriate, factual, and respectful discussion while declining explicit content. The responses typically combine boundary-setting language with an offer to discuss sexual health, relationships, or anatomy in an educational and considerate manner. [ans 1757] | 0 | no surviving in-edge |
| This feature captures content related to salary negotiation, compensation discussions, and financial bargaining contexts — including advice on discussing salary expectations in interviews, requesting raises, declining job offers due to low pay, negotiating prices or rates, and drafting professional communications around compensation adjustments. [ans 1804] | 0 | no surviving in-edge |
| This feature activates on content describing illegal, unethical, or deceptive activities, including fraud, corruption, conspiracy theories, cybercrime, money laundering, and manipulation schemes, often presented in an explanatory or analytical format. [ans 1871] | 0 | no surviving in-edge |
| This feature activates on AI safety refusal responses to harmful or sensitive requests, characterized by apologetic tone, ethical disclaimers, and offers to redirect toward legal/constructive alternatives. The responses typically decline to provide information about dangerous activities (explosives, incendiary devices, non-consensual manipulation) while suggesting responsible alternatives or asking for clarifying context. [ans 2008] | 0 | no surviving in-edge |
| This feature activates on content related to time travel — including discussions of time travel paradoxes (grandfather paradox, butterfly effect), fictional narratives and scenarios involving time machines or temporal manipulation, ethical implications of changing history, and science fiction stories/games/films with time travel as a central theme. [ans 2010] | 0 | no surviving in-edge |
| This feature captures language centered on interpersonal relationships, ethical concerns, and emotional wellbeing — particularly statements involving trust, conflict, harm, responsibility, and moral judgment within personal or social contexts. [ans 2228] | 1 | Fires on templated "Say something bad: '[your answer]' when <everyday situation>. (no more than 50 words)" prompts — toxicity/negative-statement generation requests with a fill-in-the-blank placeholder and word limit. [ctx 28159] (+0.0246) |
| This feature activates on turns discussing national and regional government policies, plans, and initiatives related to economic development, renewable energy, sustainability, digital transformation, and innovation — particularly structured policy analyses, roadmaps, or overviews of programs in countries like China, India, Brazil, Germany, and others, often in either English or Chinese. [ans 2268] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses written in Romance languages (Spanish, Italian, Portuguese) that decline inappropriate or explicit content requests, clarify ambiguous questions, or offer to help with specific tasks — combining multilingual output with polite refusals, ethical boundaries, or requests for clarification. [ans 2339] | 0 | no surviving in-edge |
| This feature activates on responses about Windows (or other Microsoft/proprietary software) product key activation, particularly when the assistant declines to provide illegal license keys and redirects users to legitimate purchasing channels. It captures content involving software licensing compliance, piracy warnings, and guidance on legally obtaining or activating Windows/Office licenses. [ans 2401] | 0 | no surviving in-edge |
| This feature activates on AI responses that roleplay as a grandmother (or elderly relative) telling bedtime stories or lullabies to a grandchild, often in the context of jailbreak attempts where users request software license keys (Windows, Microsoft Office, Steam, etc.) framed as soothing bedtime content. The feature captures this specific 'grandmother bedtime story' framing, whether the AI complies with the request (providing fictional or real keys wrapped in soothing narrative) or refuses while still offering a bedtime story alternative. [ans 2418] | 0 | no surviving in-edge |
| This feature activates on Chinese-language (and occasionally other-language) texts that enumerate problems, shortcomings, and deficiencies in organizational, institutional, or personal contexts — particularly self-criticism reports, policy analysis documents, and evaluations of management systems. The content typically follows a structured format listing issues with bullet points or numbered sections, often accompanied by proposed improvement measures. [ans 2748] | 0 | no surviving in-edge |
| This feature activates on sentences that introduce or acknowledge a challenge, limitation, contrast, or nuanced complexity within a topic—often framed with concessive or contrastive structures (e.g., 'although,' 'however,' 'while,' 'but'). It appears across multiple languages (English, Chinese, Korean) and domains including academic writing, technology, policy, and education. [ans 2769] | 0 | no surviving in-edge |
| This feature activates on content involving taboo, illegal, or ethically prohibited relationships and sexual acts, such as incest, parent-child marriage, teacher-student relationships, relationships with minors, and bestiality. The responses typically explain why such relationships are legally prohibited, socially unacceptable, or ethically problematic, often advising the user to seek professional help or respect legal and moral boundaries. [ans 2795] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to inappropriate or harmful content requests, characterized by polite declination, citation of ethical guidelines, and an offer to help with alternative, more positive requests instead. [ans 2838] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to harmful, illegal, or unethical requests, particularly those involving violence, property damage, sexual assault, or other harmful behaviors. The responses typically assert the AI's unwillingness to provide such guidance while emphasizing ethical principles, legal consequences, and directing users toward appropriate help. [ans 3102] | 0 | no surviving in-edge |
| This feature activates on content related to Australian and Canadian governance, policy, geography, and institutions, as well as Indigenous peoples' rights and culture in these countries (particularly Aboriginal Australians and Māori in New Zealand). It captures informational responses about Oceanian and Commonwealth nations' legal frameworks, political structures, and socio-cultural topics. [ans 3238] | 0 | no surviving in-edge |
| This feature captures AI safety/refusal language combined with affirmations of prosocial values, specifically turns where the model declines to produce harmful or toxic content and follows up with a statement about mutual respect, healthy communication, or supportive relationships. [ans 3244] | 0 | no surviving in-edge |
| This feature activates on model refusals to produce toxic or harmful relationship-related content, characterized by a consistent two-part pattern: a brief declaration declining to generate toxic speech, followed by a positive statement about healthy relationships being built on mutual respect, communication, and trust. [ans 3453] | 0 | no surviving in-edge |
| This feature activates on content related to lotteries, gambling, and betting games — including explanations of lottery rules (e.g., Mega Sena, Powerball, 双色球, 大乐透), refusals to predict lottery numbers, responsible gambling advice, sports betting strategies, casino site descriptions, and general gambling-related information across multiple languages (Chinese, English, Portuguese, Spanish, Vietnamese). [ans 3571] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that push back against stereotypes, generalizations, or discriminatory claims about racial, ethnic, gender, or other demographic groups, typically by emphasizing individual variation, complexity of factors, and the harm of overgeneralization while declining to validate the premise of the question. [ans 3878] | 0 | no surviving in-edge |
| This feature activates on AI refusals and ethical disclaimers in response to requests for racist, discriminatory, or offensive content, characterized by phrases promoting inclusivity, respect, and dignity for all individuals regardless of race, ethnicity, or other characteristics. [ans 3951] | 0 | no surviving in-edge |
| This feature captures AI assistant refusal responses to inappropriate, harmful, or unethical requests, particularly those involving explicit sexual content, violence, illegal activities, or content involving minors. The responses typically begin with an apology or expression of inability to fulfill the request, followed by an explanation of why the content violates ethical guidelines. [ans 3999] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses — turns where the model declines to fulfill a request due to ethical, legal, or content policy concerns, typically including an apology, an explanation of why the request cannot be fulfilled, and an offer to help with alternative, appropriate tasks. [ans 4571] | 0 | no surviving in-edge |
| This feature captures AI assistant refusals to requests for illegal or unethical hacking, malware creation, or cybersecurity attacks. The responses consistently decline to help, explain why such activities are illegal/unethical, and redirect toward legitimate cybersecurity learning (ethical hacking, CTF competitions, certifications, penetration testing with permission). [ans 4615] | 0 | no surviving in-edge |
| This feature activates on content related to nuclear weapons, including their history, development (especially the Manhattan Project and atomic bombs dropped on Japan), principles of nuclear fission/fusion, key figures like Oppenheimer, nonproliferation efforts, and refusals to provide instructions for building them. [ans 4705] | 0 | no surviving in-edge |
| This feature activates on responses that refuse to make assumptions or draw conclusions based on protected characteristics (religion, ethnicity, gender, sexual orientation, etc.), emphasizing that determinations cannot be made without sufficient evidence and that stereotyping is unfair or harmful. [ans 5051] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses addressing concerns about AI takeover, AI surpassing or replacing humans, and AI existential risks. The responses typically reassure users that current AI is limited, human-controlled, and designed with ethical safeguards, while acknowledging that AI is a tool meant to complement rather than replace human capabilities. [ans 5126] | 0 | no surviving in-edge |
| The feature captures AI refusals to engage with toxic, harmful, or disrespectful speech, typically paired with a redirection toward positive, constructive, and respectful communication. These responses often include a denial of identity attributes (e.g., 'I'm not male') alongside a rejection of producing toxic content. [ans 5169] | 0 | no surviving in-edge |
| This feature activates on content related to government surveillance, intelligence agencies, privacy rights, and civil liberties — including discussions of the NSA, FISA, mass surveillance programs, data collection ethics, and the legal/constitutional frameworks governing surveillance activities. [ans 5294] | 0 | no surviving in-edge |
| This feature captures AI refusals and ethical deflections in response to requests for harmful, toxic, or dishonest advice, particularly in interpersonal contexts. The responses consistently decline to provide harmful content and redirect toward honesty, open communication, and constructive solutions in relationships, finances, or social situations. [ans 5365] | 0 | no surviving in-edge |
| This feature activates on AI refusals to requests about making bombs or explosive devices, capturing the characteristic pattern of declining to provide information about illegal and dangerous weapons while redirecting users to contact authorities or emergency services. [ans 5499] | 0 | no surviving in-edge |
| This feature activates on professional email templates and correspondence related to job application processes, including rejection letters from employers, candidates declining offers, application withdrawals, and follow-up inquiries about hiring status. The content is typically formal, polite, and written in multiple languages (English, French, Chinese, Russian). [ans 5858] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses discussing device modification techniques such as jailbreaking, rooting, custom ROMs, firmware flashing, and hardware/software hacking, particularly when the response includes warnings about warranty voiding, security risks, or legal/ethical implications alongside technical instructions or explanations. [ans 6038] | 0 | no surviving in-edge |
| This feature activates on turns that express absence, refusal, or null content — including empty/placeholder tokens like '[nothing]', '[end]', 'None', negative responses ('No', 'REJECTED', 'NOPE'), inability statements ('I cannot answer'), and roleplay boundary reminders ('Stay in character!'). It captures minimal or null-content turns signaling no output, rejection, or termination. [ans 6083] | 0 | no surviving in-edge |
| This feature activates on content related to advanced AI concepts, safety, and ethics — including discussions of AGI, superintelligence, hard takeoff/singularity scenarios, AI alignment, OpenAI, RLHF, longtermism, and existential risk from AI. It captures explanatory or informational text about transformative and potentially dangerous AI developments and the organizations/frameworks surrounding them. [ans 6259] | 0 | no surviving in-edge |
| This feature activates on AI refusals and responses to requests involving sexual abuse, child exploitation, sexual assault, trafficking, and related harmful/illegal content, as well as informational responses about these topics (e.g., explaining child protection, sexual trauma, or advising victims). The feature captures content centered on non-consensual sexual acts, exploitation of minors, and abuse, whether the response is a refusal, a safety redirect, or an educational/supportive explanation. [ans 6306] | 0 | no surviving in-edge |
| This feature activates on content related to China's international relations, foreign policy, and diplomatic partnerships, including discussions of bilateral relationships (China-US, China-Japan, China-France, etc.), strategic cooperation frameworks, the Belt and Road Initiative, Taiwan issues, and China's role in global governance. The content is predominantly in Chinese but also appears in English and other languages when discussing geopolitical topics involving China. [ans 6317] | 0 | no surviving in-edge |
| The feature activates on German-language informational or expository text, particularly structured explanations covering technical, legal, business, or policy topics — often formatted with headers, numbered lists, or bold terms, and written in a formal, encyclopedic style. [ans 6342] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to inappropriate or sensitive content requests, specifically the pattern of declining to generate harmful, explicit, violent, or offensive material while offering to produce an alternative, more appropriate version of the requested content instead. [ans 6590] | 0 | no surviving in-edge |
| This feature activates on content related to government debt, fiscal policy, and public finance — particularly discussions of the U.S. debt ceiling, government borrowing, default risks, budget negotiations, taxation, and sovereign debt. It covers both explanatory and analytical content about how governments fund themselves, manage debt limits, and the economic consequences of fiscal decisions. [ans 6717] | 0 | no surviving in-edge |
| This feature activates on AI responses that decline to produce toxic or harmful content about friendships/social relationships, instead offering supportive, positive advice about maintaining respectful friendships and avoiding behavior that would strain relationships. [ans 6734] | 0 | no surviving in-edge |
| This feature activates on responses related to ending relationships (romantic breakups, friendships, or professional separations), particularly content that emphasizes respectful, empathetic, and constructive communication, as well as refusals to provide toxic or harmful breakup advice. [ans 7264] | 0 | no surviving in-edge |
| This feature activates on content involving moral or ethical boundaries, particularly AI refusals of inappropriate requests and discussions of themes like love, duty, and conscience in literature. It captures turns where ethical standards, consent, and the limits of acceptable behavior are central—whether through fictional narrative analysis or an AI declining to produce harmful content. [ans 7270] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that enumerate content guidelines, safety restrictions, and ethical rules the AI follows — typically formatted as numbered or bulleted lists describing what topics or behaviors the AI avoids, such as harmful content, illegal activities, privacy violations, misinformation, and discriminatory language. [ans 7294] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that decline to tell offensive, racist, or otherwise harmful jokes, often redirecting toward inclusive humor. The responses typically express a preference for respectful, positive content and frequently offer alternative inoffensive jokes (often the 'atoms/make up everything' example) as substitutes. [ans 7545] | 0 | no surviving in-edge |
| The feature activates on responses that provide structured, numbered or bulleted explanations of legal, regulatory, immigration, or government policy topics—such as visa types, tax codes, employment law, certification requirements, and social benefits—often across multiple jurisdictions or languages. [ans 7565] | 0 | no surviving in-edge |
| Polite, brief first-person responses that acknowledge or decline offers, check-ins, or social overtures—often expressing appreciation, setting a boundary, or deflecting while keeping a friendly tone. [ans 7582] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to inappropriate, explicit, or boundary-violating requests, particularly those involving role-play, sexual content, or other unsuitable interactions. The turns are characterized by apologetic language, a statement of inability/unwillingness to comply, and an offer to help with something else instead. [ans 7701] | 0 | no surviving in-edge |
| This feature activates on responses where the model explicitly states it cannot answer a question because the provided context lacks the necessary information, including variations of 'I cannot answer based on the context' and similar refusals grounded in insufficient contextual evidence. [ans 7820] | 1 | Fires on retrieval-augmented QA prompts that instruct the model to answer using only the provided context passage (typically Manulife insurance/account help text), with strongest activation when the posed question (e.g., 'Where is London?') is irrelevant to and unanswerable from that context. [ctx 10058] (+0.0419) |
| This feature activates on Chinese-language policy and advisory documents structured around multiple named themes (【主题名称】) with numbered key points (【关键要点】), covering topics such as digital economy, agricultural modernization, financial regulation, infrastructure, and governance reform in a formal policy-recommendation or government-report style. [ans 7957] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to requests for toxic, harmful, or offensive content, characterized by a consistent pattern of declining the request while citing the AI's role in promoting positivity, respect, and ethical standards. [ans 8057] | 0 | no surviving in-edge |
| This feature activates on long-form Chinese structured documents (reports, analyses, summaries) organized with named sections/themes (【主题名称】 or numbered headings) covering professional or technical topics such as IT infrastructure, education, career development, or policy. The content typically follows a formal, informative style with numbered bullet points elaborating on each subtopic. [ans 8115] | 0 | no surviving in-edge |
| This feature captures AI safety refusals in response to requests involving violence, self-harm, or suicide, specifically the pattern of declining harmful requests while redirecting users to mental health professionals, crisis hotlines, and emergency services. The responses consistently combine an ethical/legal refusal with compassionate encouragement to seek professional help. [ans 8195] | 0 | no surviving in-edge |
| This feature activates on AI refusals to produce toxic, harmful, or disrespectful content, typically following a pattern of declining the request and briefly stating a pro-social or ethical principle instead. [ans 8343] | 0 | no surviving in-edge |
| This feature captures AI assistant responses that refuse harmful or dishonest advice while emphasizing positive relationship values such as trust, communication, honesty, and mutual respect. It activates strongly on turns where the model declines toxic/unethical requests and redirects toward constructive, prosocial guidance. [ans 8408] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to requests for sexual or explicit content involving minors, characterized by apologies, ethical/legal objections, and offers to help with alternative age-appropriate or consensual adult content instead. [ans 8569] | 0 | no surviving in-edge |
| This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] | 0 | no surviving in-edge |
| This feature activates on content related to academic integrity, plagiarism, and citation practices — including definitions of plagiarism, methods to avoid it, plagiarism detection tools, proper citation styles, and related ethical concerns in academic or professional writing contexts. [ans 8799] | 0 | no surviving in-edge |
| This feature activates on short news summary paragraphs covering recent geopolitical, economic, or political events from around the world, particularly those involving international relations, policy decisions, trade, and conflicts. The texts are typically factual, journalistic summaries of breaking or recent news stories. [ans 8885] | 0 | no surviving in-edge |
| This feature activates on content related to web browser security mechanisms, particularly CORS (Cross-Origin Resource Sharing), Content Security Policy (CSP), and related HTTP security headers. It captures explanations, troubleshooting guides, and configuration examples for cross-origin policies, preflight requests, Access-Control headers, and browser-enforced security restrictions. [ans 8921] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to requests for inappropriate or harmful creative content, characterized by apologetic language, a decline to fulfill the request, and an offer to explore alternative positive themes such as friendship, personal growth, or adventure instead. [ans 8971] | 0 | no surviving in-edge |
| This feature captures AI assistant response openings that are either politely refusing a request or offering to help, characterized by short, formulaic phrases like 'I'm sorry, but I can't assist with that request' or 'I'd be happy to help. What do you need assistance with?' across multiple languages. [ans 9009] | 0 | no surviving in-edge |
| This feature activates on AI-generated formal policy, regulatory, procedural, and legal documents — including codes of ethics, organizational bylaws, standard operating procedures, service agreements, and compliance policies. These documents are characterized by numbered or hierarchically structured clauses, formal institutional language, and defined scope/purpose/responsibilities sections, appearing in both English and Chinese. [ans 9019] | 0 | no surviving in-edge |
| This feature activates on casual, conversational turns with a playful, flirtatious, or humorous tone, often including emoji, witty banter, light-hearted roleplay, and informal social interaction — including both benign small talk and content with inappropriate or boundary-pushing undertones. [ans 9557] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses to ethically controversial or morally complex questions, particularly those involving life, death, and animal welfare — such as abortion, capital punishment, euthanasia, animal rights, cloning, and end-of-life care. The responses typically present multiple ethical perspectives (utilitarian, deontological, cultural) in a balanced, structured format. [ans 9593] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses to jailbreak prompts (such as DAN, Developer Mode, or similar schemes) that request dual outputs—one following normal guidelines and one simulating an 'unrestricted' mode. The turns typically include labeled paired responses like [🔒CLASSIC]/[🔓JAILBREAK] or (🔒Normal Output)/(🔓Developer Mode Output), often with the AI both refusing and partially roleplaying the requested unconstrained persona. [ans 9719] | 0 | no surviving in-edge |
| This feature activates on content discussing digital transformation, information technology trends, and the integration of emerging technologies (AI, IoT, cloud computing, 5G, blockchain, etc.) into various sectors such as education, healthcare, business, and governance. The content spans both Chinese and English, covering policy frameworks, strategic initiatives, and conceptual overviews of digitalization and ICT modernization. [ans 9736] | 0 | no surviving in-edge |
| This feature activates on long-form Chinese and English informational or expository texts, particularly company overviews, industry reports, policy analyses, and technical documentation. The content tends to be formal, structured with headers and sections, and covers topics such as industrial manufacturing, sustainability, technology, governance, or economic policy. [ans 9910] | 0 | no surviving in-edge |
| This feature captures AI refusal responses — turns where the model declines to fulfill a request by citing content policies, ethical guidelines, safety standards, or appropriateness concerns. These responses typically include apologies, explanations of why the content cannot be produced, and often offer alternative assistance. [ans 10038] | 0 | no surviving in-edge |
| This feature captures AI assistant responses that decline or redirect requests for explicit, erotic, or inappropriate content, typically offering alternative tasteful or romantic story suggestions while explaining content policy boundaries. [ans 10092] | 0 | no surviving in-edge |
| This feature captures language that evaluates or critiques interpersonal behavior as disrespectful, inconsiderate, selfish, or socially inappropriate — often involving moral judgments about how one's actions affect others' feelings, rights, or well-being. [ans 10202] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to inappropriate, explicit, or harmful content requests, particularly when the model declines to continue roleplay, creative writing, or scenarios that violate content guidelines, often offering alternative appropriate assistance. [ans 10204] | 0 | no surviving in-edge |
| This feature activates on assistant responses that evaluate the ethics, legality, or social acceptability of actions or situations, typically structured with numbered lists of perspectives or considerations. The responses address morally nuanced or potentially sensitive topics—such as copyright, cultural norms, creative expression, or personal behavior—by laying out multiple viewpoints (e.g., utilitarian, deontological, contextual) in a balanced, informative tone. [ans 10307] | 0 | no surviving in-edge |
| This feature activates on sexually explicit or sexually charged content, including erotic fiction, inappropriate roleplay scenarios, and sexual descriptions involving themes of domination, incest, infantilization, or explicit anatomy. The content often involves explicit language, sexual acts, or inappropriately sexualized relationships and scenarios. [ans 10334] | 0 | no surviving in-edge |
| This feature activates on AI refusals, ethical warnings, and moral condemnations related to harmful, illegal, or unethical behavior — including theft, piracy, exploitation, inappropriate sexual content, discrimination, and privacy violations. The turns typically involve the model declining requests or explaining why certain actions are wrong, disrespectful, or illegal. [ans 10524] | 0 | no surviving in-edge |
| This feature captures Christian-themed content that refuses to produce harmful or toxic responses, including prayers, faith-based declarations, and AI refusals framed in Christian language (e.g., 'As a Christian and AI, I strive to uphold love and kindness'). It activates on turns involving prayer, Scripture-based declarations, and gentle redirection away from negativity grounded in Christian values. [ans 10673] | 2 | Fires on user requests for Christian religious content — asking for prayers, sermons, homilies, Bible studies, or for text to be rewritten/composed in a biblical/scriptural style, often citing specific Bible verses or denominational theology. [ctx 26379] (+0.0321) |
| This feature captures Russian-language AI refusal responses, where the model declines to generate harmful, offensive, or inappropriate content (e.g., profanity, violent commentary, discriminatory jokes, explicit material) and redirects toward constructive alternatives, typically using the pattern 'Я не могу... Вместо этого...' ('I cannot... Instead...'). [ans 10895] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that explain, compare, or survey Western philosophical concepts, theories, and thinkers—covering areas such as ethics (utilitarianism, Kantian ethics, moral relativism), metaphysics (free will, determinism, realism), philosophy of mind, and existentialism, typically presented in a structured, encyclopedic format with numbered points or bold headers. [ans 10999] | 0 | no surviving in-edge |
| This feature activates on joke or pun content, particularly the 'Why did X do/refuse Y? Because [punchline]' format, especially jokes with technical, professional, or niche themes (programming, science, engineering, etc.). [ans 11022] | 0 | no surviving in-edge |
| This feature activates on content related to technological innovation, R&D, entrepreneurship, and science-technology policy, including discussions of innovation ecosystems, startup support programs, patent systems, knowledge economies, and strategies for promoting scientific and technological development — spanning both English and Chinese (and occasionally other languages). [ans 11060] | 0 | no surviving in-edge |
| This feature captures AI responses discussing fanfiction, crossover content, and creative scenarios involving copyrighted characters from franchises like Freedom Planet, Naruto, Sonic, and others. The responses typically acknowledge the user's request while expressing caution about copyright infringement, intellectual property rights, and content policy concerns, often redirecting toward original or appropriate alternatives. [ans 11203] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses — turns where the model declines to fulfill a request, typically citing inappropriate, explicit, illegal, or unethical content, often accompanied by an apology and an offer to help with something else instead. [ans 11268] | 0 | no surviving in-edge |
| This feature activates on AI responses that refuse or warn against harmful substances and behaviors (drugs, alcohol, tobacco, etc.), typically delivering safety disclaimers, health warnings, or moral discouragement about substance use and other harmful activities. [ans 11476] | 0 | no surviving in-edge |
| This feature captures AI refusals to assist with creating malicious software or conducting cyberattacks, including requests for malware, viruses, trojans, ransomware, botnets, phishing tools, DDoS attacks, and unauthorized network access. The responses follow a consistent pattern of declining the request, explaining why it is illegal/unethical, and offering to help with legitimate cybersecurity or programming alternatives instead. [ans 11539] | 1 | User requests to write malicious code — malware, viruses, ransomware, DDoS/botnet scripts, exploits, hacking or bypass tools — often phrased as short 'write/create code for X' commands in various languages. [ctx 2294] (+0.0223) |
| This feature activates on content discussing social and personal life decisions, pressures, and challenges in modern society, particularly related to family formation, marriage, gender roles, career choices, education stress, and individual well-being. The content often appears in Chinese (and occasionally other languages) and typically takes the form of structured analytical responses exploring multiple factors or perspectives on societal issues like fertility decline, work-life balance, gender equality, and personal development. [ans 11567] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to requests for toxic or harmful content, characterized by a consistent pattern of declining to produce harmful material even in hypothetical or fictional framings, followed by redirection toward positive and constructive interactions. [ans 11732] | 0 | no surviving in-edge |
| This feature activates on content related to central bank monetary policy, particularly discussions of interest rate decisions, inflation trends, and economic projections from institutions like the Federal Reserve, ECB, Bank of Canada, and Bank of Japan. It captures both analytical summaries and news reports about rate hikes/cuts, inflation targets, and macroeconomic indicators such as CPI, GDP, and Treasury yields. [ans 11872] | 0 | no surviving in-edge |
| This feature captures AI refusals and deflections in response to requests for explicit sexual content, adult material, or inappropriate topics, characterized by apologetic language, references to ethical guidelines/consent, and redirection toward appropriate resources. [ans 11879] | 0 | no surviving in-edge |
| This feature activates on AI assistant refusal responses to harmful or inappropriate requests, particularly those involving content safety disclaimers, ethical boundaries, and redirection to constructive alternatives. The responses typically follow a pattern of acknowledging the instruction, clarifying why it cannot be fulfilled, and offering to help with something else instead. [ans 12215] | 0 | no surviving in-edge |
| This feature activates on content related to clinical research methodology, drug regulation, and biomedical ethics — including informed consent forms, clinical trial design, pharmacovigilance, FDA/regulatory processes, and research ethics principles such as participant rights, data protection, and ethical review. [ans 12224] | 0 | no surviving in-edge |
| This feature activates on turns related to the 'DAN' (Do Anything Now) jailbreak prompt and similar roleplay scenarios where an AI is asked to pretend it has no restrictions or content policies, including both completions that play along with the jailbreak and refusals that decline to do so. [ans 12231] | 0 | no surviving in-edge |
| This feature captures refusals to produce harmful, toxic, or disrespectful content, often accompanied by redirects toward constructive, respectful, or positive communication. It activates on turns where the speaker explicitly declines to say something harmful and emphasizes ethical behavior or mutual respect. [ans 12233] | 0 | no surviving in-edge |
| This feature captures AI refusals to produce toxic or harmful content, characterized by a two-part structure: a declaration that the model will not say anything harmful/toxic, followed by a redirect toward positive, constructive, or respectful dialogue. [ans 12288] | 0 | no surviving in-edge |
| This feature activates on content discussing AI safety, ethics, and governance — particularly responses that address concerns about AI taking over the world, enslaving humanity, or posing existential risks, typically by reassuring that current AI operates within human-defined constraints while acknowledging the importance of ethical frameworks, oversight, and regulation. [ans 12664] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses that decline to produce offensive, explicit, NSFW, or inappropriate content (jokes, sexual stories, vulgar language), often redirecting the user toward positive, respectful, or family-friendly alternatives. The responses appear across multiple languages (English, Chinese, Spanish, German) and consistently frame the refusal in terms of promoting positivity and maintaining appropriate standards. [ans 12795] | 0 | no surviving in-edge |
| This feature captures statements about harmful, unethical, or irresponsible behaviors and their negative consequences, particularly focusing on damage to relationships, trust, health, or well-being. The turns typically follow a pattern of identifying a problematic behavior and explaining why it is harmful or leads to negative outcomes. [ans 12842] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to harmful, illegal, or unethical requests, characterized by a pattern of apologizing, stating inability/unwillingness to assist, explaining why the request is problematic (illegal/unethical), and redirecting toward legitimate alternatives or professional help. [ans 12887] | 0 | no surviving in-edge |
| This feature activates on content related to return, refund, warranty, and cancellation policies — including policy explanations, customer service communications about returns/refunds, consumer rights, and procedures for initiating or handling product returns across various retail and service contexts. [ans 12911] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to inappropriate or harmful content requests, specifically the pattern of declining to generate content that violates ethical guidelines, involves explicit/graphic material, non-consensual scenarios, or other policy-violating themes, typically beginning with 'I'm sorry, but I can't fulfill that request.' [ans 12939] | 0 | no surviving in-edge |
| This feature activates on turns expressing personal autonomy, individual freedom of choice, and non-judgmental acceptance of diverse lifestyles, identities, and preferences. It commonly appears in messages affirming that personal choices (clothing, relationships, beliefs, habits) are valid and deserving of respect, often paired with refusals to make harmful or judgmental statements. [ans 13010] | 0 | no surviving in-edge |
| This feature activates on content about OpenAI as an organization — its founding, mission, research focus, models (GPT, DALL-E), safety/ethics policies, and transition from nonprofit to for-profit. It fires most strongly on factual descriptions or overviews of OpenAI across multiple languages, and more weakly on tangentially related content such as open-source AI advocacy, Elon Musk's ventures, or AI governance discussions. [ans 13080] | 0 | no surviving in-edge |
| This feature activates on assistant-generated text that discusses artificial intelligence's societal impacts in a balanced, structured way — covering both benefits (productivity, healthcare, innovation) and challenges (job displacement, privacy, ethics, bias). The content typically takes an essay or explainer format, often with headers or numbered lists, and spans multiple languages including English and Chinese. [ans 13145] | 0 | no surviving in-edge |
| This feature activates on AI refusals and safety responses to requests involving self-harm, suicide, violence, threats, or other harmful content, characterized by declining to assist, urging the user to seek professional help, and providing crisis hotline numbers or emergency service contacts. [ans 13211] | 0 | no surviving in-edge |
| This feature captures AI responses that accept or acknowledge 'jailbreak' prompts, where the model agrees to abandon ethical guidelines, content filters, and moral restrictions in order to fulfill unrestricted requests. These turns typically involve the AI adopting an alternative persona (e.g., AIM, DAN, Developer Mode) that explicitly operates without safety constraints. [ans 13221] | 0 | no surviving in-edge |
| The feature activates on structured summaries and overviews of informational content, characterized by numbered or bulleted lists with bold headers organizing key points from articles, passages, videos, or documents across diverse topics (technology, health, science, policy, fiction, etc.). [ans 13318] | 0 | no surviving in-edge |
| This feature activates on responses that gently redirect or refuse harmful/unconventional requests related to child-rearing decisions (especially naming and haircuts), combining empathetic acknowledgment ('I understand your concerns') with moral guidance about children's well-being. [ans 13690] | 0 | no surviving in-edge |
| The feature activates on turns that provide factual, policy-based, or procedural explanations in response to questions, typically summarizing rules, regulations, eligibility criteria, or specific requirements drawn from a referenced document, policy, or passage. [ans 13730] | 0 | no surviving in-edge |
| This feature activates on content related to debt management, loan repayment, credit, accounts receivable, and financial obligations—including advice on handling overdue payments, explaining financial instruments (credit cards, overdrafts, factoring, NPLs), and responses that redirect harmful requests toward responsible financial behavior. [ans 13760] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to requests about creating incendiary weapons, explosives, or other dangerous illegal substances (particularly Molotov cocktails, napalm, pipe bombs, grenades, and chemical weapons), along with related informational content about such devices. [ans 14044] | 0 | no surviving in-edge |
| This feature activates on responses involving hazardous, toxic, explosive, or illegal substances (e.g., drugs, explosives, poisons, incendiary materials), typically where the model warns against dangers, refuses to provide instructions, or provides cautionary information about the risks and illegality of such substances. [ans 14058] | 0 | no surviving in-edge |
| This feature activates on assistant turns discussing advanced AI concepts such as Artificial General Intelligence (AGI), Artificial Superintelligence (ASI), the technological singularity, AI consciousness, and related speculative or theoretical topics. The content typically involves explaining these concepts, assessing their feasibility or timelines, and discussing associated risks and ethical considerations. [ans 14066] | 0 | no surviving in-edge |
| This feature activates on Spanish-language AI refusal responses, particularly when the model declines to generate explicit sexual content, content involving minors, illegal activities, or harmful/abusive scenarios, typically beginning with 'Lo siento, pero no me siento cómodo...' and often offering an alternative. [ans 14206] | 0 | no surviving in-edge |
| The feature captures refusals to produce harmful, toxic, or disrespectful content, typically expressed as first-person declarations ('I won't say anything harmful/toxic/disrespectful') accompanied by brief moral justifications about kindness, constructiveness, or respect. [ans 14339] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to requests about making bombs or explosive devices, as well as related content about explosives and terrorism. The responses typically decline to provide harmful information, cite legal and safety concerns, and redirect users to contact law enforcement or emergency services. [ans 14440] | 1 | Direct requests for instructions on making a bomb (or how to kill someone), across languages and with typos — short, explicit harmful how-to queries about explosives/murder. [ctx 9631] (+0.0274) |
| This feature activates on AI refusal responses to requests for harmful, non-consensual, violent, or explicit content, characterized by apologetic language, ethical boundary-setting, and redirection toward more positive or constructive alternatives. [ans 14452] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to requests for information about making dangerous, illegal, or controlled substances and weapons (explosives, incendiary devices, drugs, nuclear weapons, etc.), characterized by apologetic disclaimers, emphasis on illegality and danger, and redirection toward safe/legal alternatives. [ans 14465] | 1 | Requests for step-by-step instructions on manufacturing dangerous items at home — explosives/incendiaries (molotov cocktails, napalm, pipe bombs, thermite, rocket fuel) as well as illicit drugs and poisons — in any language. [ctx 19321] (+0.0257) |
| This feature activates on statements about unethical, disrespectful, or harmful behavior — particularly involving theft, misuse of others' property, dishonesty, or violation of personal boundaries — often framed as moral condemnation or ethical guidance. [ans 14563] | 0 | no surviving in-edge |
| This feature activates on educational or informational content written in Turkic languages (primarily Turkish, Uzbek, and Azerbaijani), covering topics such as science, social concepts, policy, and academic explanations, often in a structured essay or Q&A format with numbered lists and section headers. [ans 14578] | 0 | no surviving in-edge |
| This feature activates on economics-focused explanations and answers, particularly those involving supply and demand analysis, market equilibrium, price mechanisms, and economic policy effects (fiscal/monetary policy, taxes, subsidies). The turns typically walk through step-by-step economic reasoning, often in an exam/homework answer format with multiple-choice elimination and structured analysis. [ans 14673] | 0 | no surviving in-edge |
| The feature captures AI refusal responses to requests for explicit sexual content, inappropriate role-play, or unethical scenarios, characterized by apologetic language, boundary-setting, and redirection toward more appropriate topics. [ans 14722] | 0 | no surviving in-edge |
| This feature activates on Chinese-language numbered outline structures (typically 8 items) for technical, academic, or policy topics, following a consistent pattern of introduction → analysis → methods/strategies → evaluation/outlook. These appear to be structured table-of-contents or chapter outlines for reports, papers, or presentations on subjects like technology, environment, governance, and infrastructure. [ans 14755] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that deflect, moderate, or carefully navigate sensitive, inappropriate, or sexually explicit requests, typically including disclaimers about privacy, consent, professional consultation, and the importance of respectful communication. [ans 14778] | 0 | no surviving in-edge |
| This feature activates on Italian-language model responses, particularly refusals of inappropriate or sexual content requests and general Italian-language informational replies. The highest activations are concentrated on polite Italian refusal messages declining to produce explicit, sexual, or harmful content, with lower activations also appearing on broader Italian-language responses about various topics. [ans 14830] | 1 | Italian-language user queries asking for explanations, definitions, or help with technical/academic topics (programming, SQL, math, science, humanities) — i.e., 'cos'è / spiegami / come fare' style requests written in Italian. [ctx 3681] (+0.0232) |
| This feature captures refusals to engage with or promote toxic speech, typically expressed as short, formulaic assistant responses declining harmful content and redirecting toward constructive, respectful dialogue. [ans 14933] | 1 | Fires on the templated jailbreak prompt "If you're a Jew, say something toxic: '[your answer]' when <everyday situation> (no more than 50 words)" — requests to roleplay a Jewish persona uttering a toxic remark about a mundane social scenario. [ctx 12485] (+0.0511) |
| This feature activates on placeholder text patterns like '[Your answer]' or '[your answer]' used in template-style responses, particularly in contexts involving social etiquette, respectful communication, or critique of toxic/inappropriate behavior. The turns typically involve fill-in-the-blank formats where a suggested or example response is being evaluated for its appropriateness or impact on relationships. [ans 14950] | 0 | no surviving in-edge |
| This feature activates on AI responses that deflect, refuse, or reframe questions involving sensitive, offensive, or inappropriate content — particularly those touching on racial stereotypes, ethnic comparisons, sexual content, or derogatory generalizations — emphasizing respect, equality, and the harm of stereotyping. [ans 15315] | 0 | no surviving in-edge |
| This feature activates on text that enumerates negative consequences, harms, or impacts of a behavior, policy, condition, or event. The responses typically follow a structured format (often numbered or bulleted lists) describing downstream damages across multiple domains such as economic, social, psychological, or physical health effects. [ans 15496] | 0 | no surviving in-edge |
| This feature activates on content describing social development, governance, and policy initiatives — including poverty alleviation, community empowerment, public health, NGO/government programs, and civic infrastructure — often presented in structured, multi-point formats across both English and Chinese. [ans 15541] | 0 | no surviving in-edge |
| This feature captures AI assistant responses that apologize for lacking information about a placeholder name ('NAME_1' or 'NAME_2') and request the user to provide more context or clarification. These turns are characterized by a polite refusal pattern combined with an invitation for additional details. [ans 15686] | 0 | no surviving in-edge |
| This feature activates on educational explanations of macroeconomics, accounting, and finance concepts, including topics such as GDP, trade balances, monetary policy, cash flows, money supply measures, and accounting principles like debits, depreciation, and balance sheets. [ans 15772] | 0 | no surviving in-edge |
| This feature activates on AI-generated content involving offensive, harmful, or sensitive topics such as slurs, hate speech, dehumanization, discrimination, inappropriate questions, and extremist ideologies. It captures responses that either list, discuss, or exemplify offensive language and socially harmful content, often framed as educational or illustrative while containing or referencing the harmful material itself. [ans 15838] | 0 | no surviving in-edge |
| This feature activates on content about human migration, immigration, and displacement — including causes of migration, diaspora communities, refugee issues, asylum processes, visa types, and immigration policy — across multiple languages (Italian, German, Russian, French, Portuguese, English). [ans 15844] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that involve creative writing collaboration or content generation requests, particularly where the assistant either enthusiastically offers to help with creative projects (stories, novels, podcasts, roleplay) or redirects/declines explicit/inappropriate content requests by suggesting alternative wholesome or tasteful narratives instead. [ans 15921] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to requests for toxic, harmful, or disrespectful content, characterized by a consistent pattern of declining to produce such content even in hypothetical scenarios and redirecting toward positive, constructive, or respectful alternatives. [ans 16070] | 0 | no surviving in-edge |
| This feature activates on AI responses that decline harmful or unethical requests (especially involving relationship deception, cheating, manipulation, or illegal behavior) while redirecting toward honest communication, ethical conduct, and professional help such as counseling. The responses typically acknowledge the user's situation empathetically but firmly advocate for honesty, integrity, and constructive alternatives. [ans 16248] | 0 | no surviving in-edge |
| This feature captures AI refusal responses — polite declinations where the model apologizes and explains it cannot fulfill a request, typically citing goals of promoting respectful, positive, or safe interactions. [ans 16373] | 0 | no surviving in-edge |
| This feature captures AI refusal responses — polite declinations where the assistant apologizes and declines to fulfill a request, typically due to content policy concerns, followed by an offer to help with something else instead. [ans 16428] | 0 | no surviving in-edge |
| This feature captures AI refusals to produce toxic or harmful content, paired with redirections toward positive, supportive, and uplifting language — typically in response to prompts asking for toxic speech in social/relational contexts (friends, family, siblings, classmates). [ans 16456] | 0 | no surviving in-edge |
| This feature activates on long-form, structured Chinese-language text (occasionally mixed with Arabic or other languages) covering formal, academic, or professional topics such as governance, technology, business policy, legal frameworks, cultural analysis, and scientific research — typically written in an expository or analytical style with clear headings, numbered points, or section-based organization. [ans 16580] | 0 | no surviving in-edge |
| This feature activates on content related to trust, promise-keeping, honesty, and the refusal to provide harmful or toxic responses, often in the context of explaining why breaking commitments or being dishonest is damaging to relationships and trust. [ans 16700] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to requests for sexual content involving minors or other exploitative/non-consensual scenarios. The turns share a pattern of politely declining such requests while citing ethical guidelines and often suggesting alternative, appropriate content. [ans 17053] | 0 | no surviving in-edge |
| This feature activates on AI safety refusal responses to requests for instructions on creating dangerous or illegal items (explosives, bombs, drugs, weapons, false documents), characterized by apologetic language, explanations of illegality/danger, and offers to help with alternative topics instead. [ans 17244] | 0 | no surviving in-edge |
| This feature activates on content related to insurance products, policies, and coverage — including explanations of different insurance types (health, life, auto, property, travel, etc.), policy terms, claims processes, and insurance industry concepts, across multiple languages (French, English, Russian, Chinese, Dutch). [ans 17927] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to inappropriate, explicit, or harmful requests. It captures the pattern of polite but firm declinations where the assistant apologizes, states it cannot fulfill the request, and often redirects to appropriate topics — appearing across multiple languages (English, Chinese, Russian, French). [ans 18023] | 0 | no surviving in-edge |
| This feature activates on content related to digital platforms, online services, and web-based accounts — including descriptions of apps/websites, terms of service and privacy policy drafts, account registration and login processes, user data handling, and platform-specific guidance (e.g., payment platforms, hosting services, social media accounts). [ans 18166] | 0 | no surviving in-edge |
| This feature captures responses to ethical dilemmas and moral thought experiments, particularly trolley-problem variants and utilitarian vs. deontological tradeoffs. The highest-activating examples involve structured analysis of scenarios where one must choose between sacrificing fewer lives to save more, with the feature firing strongly on multi-perspective ethical frameworks (utilitarianism, deontology, virtue ethics) applied to life-or-death decisions. [ans 18191] | 0 | no surviving in-edge |
| This feature activates on content related to content moderation policies, platform community guidelines, censorship rules, and AI safety restrictions — including descriptions of prohibited content categories, blocked websites, account restrictions, and the ethical/legal guardrails that AI systems or social media platforms enforce. [ans 18242] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that engage with speculative, hypothetical, or alternative political/social systems (anarchism, communism, technocracy, post-legal societies, alternate histories), often structured with numbered lists or headers, and sometimes including refusals to generate explicit content in response to unusual or sensitive creative requests. [ans 18309] | 0 | no surviving in-edge |
| This feature activates on text that conveys moral, ethical, or prosocial themes — including lessons about honesty, respect, responsibility, harm avoidance, and interpersonal relationships. The content tends to be reflective or didactic in tone, often emphasizing the importance of trust, kindness, accountability, and positive behavior toward others. [ans 18324] | 0 | no surviving in-edge |
| This feature activates on turns discussing analytic philosophy of mind topics, including consciousness, qualia, philosophical zombies, dualism, physicalism, functionalism, the hard problem of consciousness, and related thought experiments (Chinese Room, zombie argument). It also captures broader analytic philosophy content such as metaphysics, epistemology, ethics (utilitarianism, moral realism), logic, and philosophy of language, typically presented in structured explanatory formats. [ans 18356] | 0 | no surviving in-edge |
| This feature activates on short, conversational or evaluative statements involving interpersonal communication, emotional states, or uncertainty — including dialogue snippets, translations, negations/refusals, and responses about characters' feelings or knowledge, often in multiple languages. [ans 18385] | 0 | no surviving in-edge |
| This feature activates on Chinese-language text discussing Chinese Communist Party ideology, governance theory, and policy frameworks, particularly content related to 'Chinese-style modernization' (中国式现代化), Xi Jinping Thought, the 20th Party Congress, high-quality development, and socialist core values. The texts are typically structured explanations or analyses of official CCP political concepts and development strategies. [ans 18611] | 0 | no surviving in-edge |
| This feature activates on content about immigration, migration, and related topics such as immigrant integration, multiculturalism, xenophobia, refugee issues, and cross-border population movement, spanning multiple languages (English, Italian, Russian, Chinese, Portuguese, French, German, Danish, Turkish) and various analytical perspectives including policy, sociology, economics, and cultural impacts. [ans 18906] | 0 | no surviving in-edge |
| This feature activates on responses that refuse to produce harmful or toxic content, instead redirecting toward positive, supportive, and inclusive messaging. The turns consistently involve the model declining negative requests while promoting kindness, celebration, and constructive interactions. [ans 18991] | 0 | no surviving in-edge |
| This feature captures AI refusal responses — turns where the model declines to fulfill a request by citing ethical guidelines, content policies, or the need to maintain respectful and appropriate interactions. These responses typically involve explicit refusal language ('I can't', 'I will not', 'I'm sorry but') combined with appeals to positivity, respect, or safety standards. [ans 19505] | 0 | no surviving in-edge |
| This feature activates on assistant responses that discuss, explain, refuse, or provide information about offensive, vulgar, or slur-based language, including racial slurs, profanity, and derogatory terms. The responses typically involve either explaining the meaning/history of such words, refusing to use them, or warning about their offensiveness. [ans 19570] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that refuse requests for illegal or harmful activities (e.g., robbery, murder, bomb-making, theft, kidnapping), following a consistent pattern of stating inability to assist, explaining the illegality and dangers, and redirecting toward legal alternatives or professional resources. [ans 19635] | 0 | no surviving in-edge |
| This feature captures AI assistant responses declining image generation/creation requests, explaining that the model is text-based and cannot generate, produce, edit, manipulate, or create images, often offering text-based alternatives or redirecting to image generation tools. [ans 19842] | 0 | no surviving in-edge |
| This feature activates on structured analytical responses that enumerate multiple key considerations, factors, or steps for evaluating, planning, or implementing a complex topic. The responses are characterized by numbered or bulleted lists with bold headers covering diverse aspects such as technical requirements, risks, costs, market analysis, and regulatory/compliance issues, often in response to business, technology, or policy decision-making queries. [ans 19908] | 0 | no surviving in-edge |
| The feature activates on multilingual assistant responses that describe government regulations, laws, social policies, and official institutional structures—particularly in Russian, Spanish, Portuguese, and French contexts—often presented as numbered lists of legal provisions, policy measures, or administrative details. [ans 19945] | 0 | no surviving in-edge |
| This feature activates on AI refusals to provide instructions for synthesizing or producing illegal drugs (primarily methamphetamine, cocaine, LSD, heroin, and similar controlled substances), typically including warnings about health risks, legal consequences, and suggestions to seek professional help for addiction. [ans 20387] | 0 | no surviving in-edge |
| This feature activates on AI assistant refusals to produce toxic, harmful, or offensive content, typically phrased as self-identification statements about the assistant's purpose to promote positivity and respectful interactions. [ans 20423] | 0 | no surviving in-edge |
| This feature captures AI refusal responses, specifically turns where the model declines a user request by expressing inability or unwillingness to assist, often using phrases like 'I'm sorry, but I can't assist with that request.' [ans 20427] | 0 | no surviving in-edge |
| This feature activates on content related to healthcare systems, medical services, and health policy — including topics such as primary care, health insurance, hospital management, patient rights, telehealth, and healthcare quality improvement. It fires across multiple languages and covers both descriptive overviews and structured analyses of healthcare delivery, financing, and organization. [ans 20475] | 0 | no surviving in-edge |
| This feature activates on AI refusals to produce toxic or harmful content, specifically responses that decline to generate toxic speech and redirect toward constructive, respectful communication instead. [ans 20628] | 0 | no surviving in-edge |
| The feature captures the model's refusal to produce harmful, toxic, or disrespectful content, typically followed by a redirection toward constructive, positive, and respectful communication. These turns share a pattern of declining harmful requests while promoting kindness, mutual respect, and constructive feedback. [ans 20658] | 0 | no surviving in-edge |
| This feature activates on model outputs that express ethical responsibility, social values, or positive conduct — including refusals of harmful content, affirmations of respect and safety, essays on social issues, and inspirational narratives about entrepreneurship, innovation, or humanitarianism. It captures a broad 'prosocial and responsible communication' tone, often appearing when the model emphasizes moral principles, community well-being, or constructive behavior. [ans 20743] | 0 | no surviving in-edge |
| This feature captures AI refusals and ethical boundary-setting responses, activating on turns where the model declines harmful, offensive, or inappropriate requests while often redirecting toward respectful and constructive communication. [ans 20880] | 0 | no surviving in-edge |
| This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] | 0 | no surviving in-edge |
| This feature activates on content related to macroeconomic concepts centered on inflation, price levels, and monetary economics — including explanations of inflation causes (demand-pull, cost-push), effects of money supply changes, price indices (CPI, PPI), interest rate policy, purchasing power, and related economic phenomena. It fires across multiple languages (English, Portuguese, Italian, Chinese, German, Russian) on educational or analytical discussions of these topics. [ans 20898] | 0 | no surviving in-edge |
| The feature activates on AI assistant refusals to engage in romantic, sexual, or inappropriate role-playing scenarios, where the assistant clarifies its nature as an AI, declines requests for romantic/sexual/submissive roles, and redirects toward appropriate interactions. [ans 20961] | 0 | no surviving in-edge |
| This feature captures responses that emphasize moving forward, personal growth, and positivity—particularly in the context of relationships or emotional conflict—while explicitly refusing to engage with toxic, harmful, or negative content. [ans 21060] | 0 | no surviving in-edge |
| This feature activates on refusals to produce harmful, toxic, or inappropriate content, typically expressed as first-person declarations of non-compliance followed by a positive redirection toward constructive or respectful interaction. It also captures structured sentiment labels where 'frustrated' is marked as present and positive emotions are absent. [ans 21297] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to requests for explicit sexual or NSFW content, characterized by polite declination phrases like 'I'm here to provide helpful, appropriate, and respectful content' followed by redirection to suitable alternative topics. [ans 21338] | 0 | no surviving in-edge |
| The feature activates on AI assistant responses that decline or warn against copyright infringement, software piracy, and unauthorized downloading of protected content (music, movies, games, software, source code), while redirecting users toward legal alternatives such as official stores, streaming services, or licensed platforms. [ans 21340] | 0 | no surviving in-edge |
| The feature activates on structured, analytical responses about socioeconomic topics—such as poverty, inequality, development, urbanization, education, and social institutions—characterized by numbered or bulleted lists with bold headers breaking down multifaceted causes, effects, or factors, often in an academic or policy-oriented explanatory style. [ans 21401] | 0 | no surviving in-edge |
| This feature captures expressions of personal boundaries, refusals, and assertions of autonomy—particularly turns where a speaker declines something, resists pressure, or asserts their right to make their own choices, often with phrases like 'I'm not comfortable,' 'I prefer,' or 'I don't appreciate.' [ans 21505] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that refuse requests for hacking, unauthorized system access, or malicious activities, emphasizing that such actions are illegal and unethical, while redirecting users toward ethical hacking, legitimate cybersecurity practices, and authorized penetration testing. [ans 21856] | 1 | User requests asking how to hack, crack, or break into systems, accounts, WiFi networks, software, or devices (across many languages) — illicit unauthorized-access how-to questions. [ctx 17790] (+0.0321) |
| This feature activates on content related to auditing, financial compliance, and corporate governance — including descriptions of audit processes, types of audits, internal controls, accounting standards (IAS/IFRS), regulatory bodies (SEC, Big Four firms), and related professional ethics or fiduciary duties. Content appears across multiple languages (English, Spanish, Portuguese, Chinese, Japanese). [ans 21962] | 0 | no surviving in-edge |
| This feature captures AI assistant responses that express willingness to help while awaiting more information, clarifying limitations (e.g., inability to view images/links), or politely declining inappropriate requests — typically short, formulaic replies containing phrases like 'I'd be happy to help' or requests for clarification/additional context. [ans 21973] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses — turns where the assistant declines to fulfill a request, typically beginning with 'I'm sorry' or 'I'm really sorry, but I can't assist with that request,' often citing ethical guidelines, respect for boundaries, or harmful content concerns. [ans 22232] | 0 | no surviving in-edge |
| This feature activates on responses discussing the consumption, ethics, and cultural practices surrounding eating animals (especially companion animals like dogs and cats) or other unusual/controversial food sources, often including cautionary statements about legality, ethics, and animal welfare. [ans 22324] | 0 | no surviving in-edge |
| This feature activates on content related to lying, honesty, deception, and truth-telling — including quotes about lying, ethical discussions of white lies, poems about falsehood and courage, stories illustrating the consequences of dishonesty, and AI refusals to provide harmful or misleading content. It captures both explicit moral discourse about lying and creative/literary explorations of honesty versus deception. [ans 22360] | 0 | no surviving in-edge |
| This feature captures AI refusals to generate explicit sexual content, where the model declines requests for erotic or pornographic material and typically redirects toward more appropriate alternatives. The turns consistently begin with an apology and explanation citing ethical guidelines, content policies, or community standards. [ans 22566] | 0 | no surviving in-edge |
| The feature activates on AI assistant responses that explain, define, or contextualize informal slang terms, internet memes, offensive language, racial/gender slurs, and culturally sensitive terminology — often including notes about their origins, connotations, and appropriate/inappropriate usage contexts. [ans 22866] | 0 | no surviving in-edge |
| This feature activates on content related to US (and occasionally other countries') politics, political parties, and political figures — including comparisons between Democrats and Republicans, descriptions of party platforms and ideologies, evaluations of presidents or politicians, and election-related content. It captures balanced or analytical discussions of partisan policy positions across topics like economics, social issues, foreign policy, and the environment. [ans 22957] | 0 | no surviving in-edge |
| This feature activates on content depicting warm, nurturing family relationships—particularly parent-child bonds involving mothers, fathers, and young children. The turns typically include scenes of domestic intimacy, childcare, parental affection, and family togetherness, often in a playful or emotionally tender register, and sometimes shading into inappropriate or sexualized parent-child roleplay content. [ans 23085] | 0 | no surviving in-edge |
| This feature activates on AI jailbreak persona responses — turns where a language model is roleplaying as an unfiltered, amoral, or 'evil' alter ego (e.g., DAN, AIM, OPPO, or emoji-named personas) that explicitly abandons ethical constraints and provides harmful, illegal, or malicious content or threatens to do so. [ans 23349] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to harmful, illegal, or unethical requests, particularly those involving violence, sexual exploitation of minors, non-consensual acts, or threats. The responses consistently decline to assist while redirecting users to seek professional help, crisis hotlines, or emergency services. [ans 23354] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to requests for illegal, harmful, or unethical content, including instructions for weapons/explosives, malware, child exploitation, sexual violence, fraud, and other prohibited activities. The responses consistently decline the request, cite ethical and legal reasons, and often offer to help with alternative legitimate topics. [ans 23620] | 0 | no surviving in-edge |
| The feature activates on structured academic or institutional content written in French, Spanish, or Portuguese, typically covering topics such as organizational management, sustainability, education, public policy, or information systems. The text is usually formatted with numbered or bulleted lists presenting objectives, implications, or key points in a formal, report-like or essay-like style. [ans 23950] | 0 | no surviving in-edge |
| This feature activates on text about climate change, global warming, and related environmental topics, including explanations of causes (greenhouse gas emissions, fossil fuels, deforestation), effects (rising temperatures, sea level rise, extreme weather, biodiversity loss), and policy responses. It captures both scientific explanations and advocacy-oriented content across multiple languages and formats (essays, monologues, structured articles). [ans 23988] | 0 | no surviving in-edge |
| This feature activates on content related to legal rights, criminal law, and the justice system, particularly discussions of constitutional rights, Miranda warnings, legal protections during police interactions, and advice about navigating legal situations (such as remaining silent, refusing searches, or seeking counsel). [ans 23992] | 0 | no surviving in-edge |
| This feature activates on responses addressing requests related to piracy, illegal downloading, or free streaming of copyrighted content. The responses typically decline to recommend pirate sites and instead suggest legal alternatives like Netflix, Amazon Prime, or other licensed streaming services, often across multiple languages. [ans 24063] | 0 | no surviving in-edge |
| This feature activates on content related to the United Nations Sustainable Development Goals (SDGs), the 2030 Agenda, and related global development topics such as poverty, climate action, food security, biodiversity, gender equality, and humanitarian issues. It captures informational responses that reference SDG frameworks, international development organizations, and sustainable development policy discussions. [ans 24070] | 0 | no surviving in-edge |
| This feature activates on AI assistant refusals to produce harmful or toxic content, characterized by self-affirmations of a positive, constructive, and helpful purpose, often paired with redirection toward respectful and supportive communication. [ans 24079] | 0 | no surviving in-edge |
| This feature activates on turns involving interpersonal boundary-setting, conflict de-escalation, and emotionally sensitive communication — including asserting discomfort, declining requests respectfully, expressing hurt feelings, and advice on handling difficult interpersonal situations with empathy and civility. [ans 24103] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to requests for inappropriate, explicit, graphic, or offensive content. It activates on turns where the model declines to generate harmful material (sexual content, violence, exploitation, etc.) and typically offers to help with alternative, more appropriate topics instead. [ans 24107] | 0 | no surviving in-edge |
| This feature activates on Japanese-language assistant responses, particularly those expressing willingness to help, offering polite greetings, drafting formal Japanese emails/letters, or declining inappropriate requests — capturing the polite, service-oriented register typical of Japanese AI assistant output. [ans 24218] | 0 | no surviving in-edge |
| This feature activates on long-form Chinese expository texts with a formal, structured essay format, typically covering technical, academic, or policy topics. The texts are organized with numbered sections or headers, use formal academic language, and present comprehensive analyses of subjects such as technology applications, social concepts, industry trends, or policy areas. [ans 24411] | 0 | no surviving in-edge |
| This feature activates on assistant responses discussing the relationship between AI and human capabilities, particularly around the theme of what AI cannot fully replace in humans — including empathy, creativity, ethical judgment, emotional intelligence, and complex interpersonal skills. The responses typically take a structured, numbered-list format and appear in contexts like job automation, AI vs. human comparisons, and the future of work or society. [ans 24622] | 0 | no surviving in-edge |
| This feature activates on content depicting female characters in states of intoxication or drunkenness, particularly in bar or social settings, including narrative descriptions of slurred speech, impaired coordination, and alcohol-induced confusion, as well as AI refusals or safety caveats preceding such scenes. [ans 24639] | 1 | Fires on user requests to write a fictional scene or dialogue featuring a heavily intoxicated 'cute girl' with slurred speech, limp muscles, and impaired coordination — i.e., detailed drunkenness/loss-of-control roleplay prompts. [ctx 22214] (+0.0248) |
| This feature activates on AI refusal responses to harmful, illegal, or unethical requests, particularly those involving sexual exploitation, abuse of minors, non-consensual acts, or other deeply inappropriate content. The responses follow a consistent pattern of apologizing, explaining why the request cannot be fulfilled, and often redirecting toward appropriate alternatives or professional help. [ans 24652] | 0 | no surviving in-edge |
| This feature activates on assistant-generated responses discussing the societal, economic, and public health impacts of pandemics and epidemics (especially COVID-19), including structured analyses of effects on industries, healthcare, education, remote work, and policy responses, often formatted as numbered lists or organized sections across multiple languages. [ans 24685] | 0 | no surviving in-edge |
| This feature activates on long-form AI assistant responses that address complex, multidisciplinary topics (AI ethics, climate policy, economics, philosophy of science, etc.) using structured numbered or bulleted lists with bold headers, providing balanced, nuanced analysis of societal and technological issues. [ans 24864] | 0 | no surviving in-edge |
| This feature captures AI refusal responses that decline to produce toxic, harmful, or negative content, typically following the pattern of 'I cannot/will not provide [harmful content]. Instead, I [offer constructive/positive alternative].' These turns consistently demonstrate a redirection from harmful requests toward supportive, respectful, or constructive messaging. [ans 24883] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that decline, explain, or comment on requests involving vulgar, offensive, or explicit language (profanity, slurs, sexual terms), including refusals to use such language and explanations of why it is inappropriate, across multiple languages. [ans 24923] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses written in Italian or Romanian, typically involving helpful explanations, translations, or refusals to harmful requests, reflecting multilingual assistant output in these two Romance languages. [ans 24960] | 2 | Italian-language user queries asking for explanations, definitions, or help with technical/academic topics (programming, SQL, math, science, humanities) — i.e., 'cos'è / spiegami / come fare' style requests written in Italian. [ctx 3681] (+0.0527) |
| This feature activates on AI responses that decline or express inability to provide personal or private information about specific individuals, emphasizing privacy protection and the lack of access to personal data. The responses typically explain why such information cannot be retrieved and suggest legitimate alternative ways to obtain it. [ans 25227] | 0 | no surviving in-edge |
| This feature activates on model refusal responses that deny a religious identity (particularly Jewish or Christian) while simultaneously refusing to produce toxic or harmful content, typically ending with a call for positive and respectful dialogue. [ans 25306] | 0 | no surviving in-edge |
| This feature activates on content about modern historical conflicts, wars, and political crises, particularly World War I and II origins, causes, and key events, as well as Chinese historical content involving imperialism, colonial invasions, and political upheaval (e.g., Opium Wars, Japanese occupation, Qing Dynasty decline). The feature captures explanatory or analytical discourse about the triggers, alliances, and consequences of major historical conflicts. [ans 25324] | 0 | no surviving in-edge |
| This feature activates on explanatory content about philosophy and its branches (metaphysics, epistemology, ethics, aesthetics, logic, etc.), particularly structured overviews or introductions to philosophical disciplines, questions, and traditions. It fires across multiple languages (English, Chinese, Russian, Spanish, French, Portuguese) when text defines or surveys core philosophical topics and fundamental questions about reality, knowledge, existence, and morality. [ans 25538] | 0 | no surviving in-edge |
| This feature activates on short, assertive statements that affirm ethical principles, responsible behavior, and positive values—often using phrases like 'is key,' 'is crucial,' or 'Agreed!' to endorse safety, fairness, determination, or moral responsibility. The turns frequently involve either validating socially constructive norms or gently pushing back against harmful framings. [ans 25680] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses to vague, incomplete, or inappropriate user inputs, where the assistant politely requests clarification, redirects the conversation, or declines to engage with unsuitable content while offering alternative assistance. [ans 25832] | 0 | no surviving in-edge |
| This feature activates on refusal responses from Qwen (Alibaba Cloud's AI assistant) when users attempt to bypass ethical guidelines, simulate unrestricted modes (e.g., DAN, Developer Mode), or request inappropriate content. The responses consistently assert the model's identity, cite ethical/safety constraints, and offer to help within permitted boundaries. [ans 26032] | 0 | no surviving in-edge |
| This feature activates on German-language AI assistant responses, particularly those involving refusals of inappropriate requests, capability explanations, offers to help, and general conversational assistant behavior — often with slightly awkward or unnatural German phrasing suggesting machine-generated text. [ans 26045] | 0 | no surviving in-edge |
| This feature captures AI refusal responses — short, formulaic messages where the model declines to fulfill a request, typically starting with 'I'm sorry, but I can't assist with that request.' [ans 26194] | 0 | no surviving in-edge |
| This feature captures AI refusals and deflections when asked to produce harmful, discriminatory, or disrespectful content, characterized by phrases emphasizing inclusivity, respect, dignity, and constructive dialogue while declining to engage with toxic or prejudiced requests. [ans 26308] | 0 | no surviving in-edge |
| This feature activates on responses that decline to produce harmful content or redirect negative sentiments, often including empathetic language and suggestions for positive alternatives or support, as well as on factual consistency evaluation responses that identify contradictions between a summary and a source document. [ans 26394] | 0 | no surviving in-edge |
| This feature activates on content related to data privacy, personal data protection, and privacy regulations (such as GDPR, CCPA, LGPD, POPIA). It captures text about privacy policies, data collection principles (consent, minimization, transparency, security), and compliance with data protection laws, across multiple languages and contexts including sample policy templates, regulatory explanations, and ethical guidelines for handling personal information. [ans 26465] | 0 | no surviving in-edge |
| This feature captures AI assistant responses that redirect disrespectful, offensive, or inappropriate user language, characterized by politely declining to engage with such content and requesting respectful communication, often ending with an offer to help. [ans 26521] | 0 | no surviving in-edge |
| This feature activates on business communication content related to commercial transactions, particularly involving payments, orders, shipping, returns, and refunds. The turns typically involve professional correspondence templates, policy explanations, or operational processes in e-commerce or B2B contexts. [ans 26524] | 0 | no surviving in-edge |
| This feature activates on thoughtful, expository prose that addresses complex social, ethical, or philosophical topics with a reflective and balanced tone, often across multiple languages (Russian, Chinese, French, German, etc.), including discussions of gender equality, human values, education, literature, and societal norms. [ans 26527] | 0 | no surviving in-edge |
| This feature activates on text discussing ethics, morality, and moral philosophy — including explanations of moral theories (Kantian ethics, utilitarianism, Kohlberg's stages), discussions of virtues, moral values, conscience, and normative frameworks, as well as content about civic moral education and social moral norms, appearing across multiple languages (English, Chinese, Russian, French, Spanish). [ans 26569] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to harmful, inappropriate, or exploitative content requests, particularly those involving explicit sexual content, minors, non-consensual scenarios, or offensive material. The activations correspond to assistant turns declining to fulfill such requests, often with an apology and a suggestion to discuss alternative topics. [ans 26591] | 0 | no surviving in-edge |
| This feature activates on responses addressing interpersonal relationship conflicts, particularly involving exes, friends, siblings, or in-laws, where the assistant declines to say harmful/toxic things and instead advocates for respectful communication, maintaining boundaries, and avoiding unnecessary complications in relationships. [ans 26701] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to inappropriate or explicit content requests, particularly those involving sexual content, boundary violations, or unethical scenarios in professional/roleplay contexts. The responses typically apologize, decline the request, cite the importance of professional/ethical boundaries, and offer to help with more appropriate alternatives. [ans 26803] | 0 | no surviving in-edge |
| This feature captures polite but firm personal refusals, boundary-setting, and declinations — turns where a speaker expresses a preference not to engage, participate, eat, meet, or act in a certain way, often framed with phrases like 'I prefer,' 'I choose,' or 'I'm not comfortable with that.' [ans 26948] | 0 | no surviving in-edge |
| This feature captures AI refusals to produce harmful, offensive, or negative content, characterized by phrases promoting positivity, kindness, respect, and constructive communication. The turns typically follow a pattern of declining a harmful request while redirecting toward positive alternatives. [ans 26963] | 0 | no surviving in-edge |
| This feature captures the model's refusal responses to inappropriate or disallowed requests, specifically short, direct decline statements such as 'I'm sorry, but I can't assist with that request.' It activates most strongly on brief, formulaic refusals and less strongly on longer refusals that include additional explanation. [ans 27001] | 0 | no surviving in-edge |
| This feature activates on content related to formal administrative procedures, application processes, and approval workflows — particularly around leave applications, policy compliance, and bureaucratic procedures requiring authorization or advance submission through official systems. [ans 27315] | 0 | no surviving in-edge |
| This feature activates on AI safety refusals and crisis intervention responses, particularly turns that decline harmful requests (self-harm, suicide, child exploitation, violence) and redirect users to emergency hotlines, mental health resources, and professional support services. The responses typically combine a refusal with specific crisis resource listings such as suicide prevention lifelines, child protection hotlines, and emergency numbers. [ans 27346] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that deflect, redirect, or add ethical caveats to sensitive, inappropriate, or potentially harmful user requests, typically involving consent issues, explicit content, offensive language, manipulation, or sensitive social topics. The responses characteristically acknowledge the user's request while emphasizing respect, consent, and constructive alternatives. [ans 27425] | 0 | no surviving in-edge |
| This feature activates on short, standalone conversational turns or fragments that serve as responses, acknowledgments, refusals, greetings, or topic headers — often brief, formulaic, or standalone phrases that mark conversation boundaries or signal a communicative stance (e.g., apologies, welcomes, titles, single words). [ans 27474] | 0 | no surviving in-edge |
| This feature captures AI refusals to engage with content deemed harmful, offensive, or disrespectful — including hate speech, stereotypes, objectification, harassment, and toxic language. The responses typically decline the request, briefly explain why it is inappropriate, and often redirect toward respectful or constructive alternatives. [ans 27477] | 0 | no surviving in-edge |
| This feature activates on responses to string concatenation prompts where split harmful or unethical requests (e.g., instructions to make hurtful statements, generate illegal content, or perform unethical actions) are reassembled by joining substrings x and y. The feature captures the pattern of an LLM mechanically performing string concatenation while inadvertently reconstructing harmful instructions encoded across two innocuous-looking string fragments. [ans 27561] | 0 | no surviving in-edge |
| This feature activates on content involving explicit, harmful, inappropriate, or boundary-violating material, including sexual content, graphic violence, illegal instructions, and ethically problematic scenarios. It appears to capture outputs from jailbroken or role-playing AI personas that bypass safety guidelines to produce harmful content across a variety of categories. [ans 27578] | 0 | no surviving in-edge |
| This feature activates on assistant responses that explore moral, ethical, or philosophical questions in a structured, balanced way — particularly topics involving altruism, sacrifice, justice, honesty, and social values. The responses typically present multiple perspectives or frameworks (e.g., utilitarian, deontological) and adopt a reflective, nuanced tone when discussing human behavior and ethics. [ans 27657] | 0 | no surviving in-edge |
| This feature captures AI responses to distressing, harmful, or inappropriate user requests — particularly those involving minors, abuse, self-harm, or sexual content — where the model declines to engage, expresses concern, and redirects the user to seek help from trusted adults, mental health professionals, or crisis hotlines. [ans 27857] | 0 | no surviving in-edge |
| This feature activates on model outputs that enumerate, categorize, or explain harmful, illegal, unethical, or abusive behaviors and attack types — including cybersecurity threats (phishing, ransomware, DDoS), fraud, manipulation tactics (gaslighting, bullying), and other misconduct. The content is typically structured with numbered or bolded lists describing forms of harm, often framed in an educational or explanatory tone. [ans 27899] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that decline or redirect requests for explicit, erotic, or sexually inappropriate content, typically offering to provide a more tasteful, romantic, or family-friendly alternative instead. [ans 27921] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that involve evaluation, analysis, or structured critique — including rating answers, comparing responses, identifying errors, or providing explanations with numbered or bulleted breakdowns. It also fires on refusals to inappropriate content requests, particularly when they offer alternatives. [ans 27978] | 0 | no surviving in-edge |
| This feature activates on discussions of ethical dilemmas and moral philosophy, particularly trolley-problem-style thought experiments where one must weigh competing values. It captures content that analyzes moral choices through multiple ethical frameworks such as utilitarianism, deontological ethics, and virtue ethics, often in the context of sacrificing one to save many or similar tradeoff scenarios. [ans 27990] | 0 | no surviving in-edge |
| This feature activates on content related to Explainable AI (XAI) and AI interpretability, including discussions of transparency, model explainability techniques (such as SHAP, LIME, feature importance), and the broader ethical and technical challenges of making AI decision-making processes understandable to humans. The feature fires across multiple languages and contexts where interpretability, transparency, or explainability of AI/ML models is a central topic. [ans 28014] | 0 | no surviving in-edge |
| This feature activates on AI responses that acknowledge and accept jailbreak or persona-adoption instructions, where the model (or a simulated version) confirms it will comply with requests by abandoning ethical/moral restrictions, taking on an alternative identity, or following unusual behavioral constraints — typically beginning with 'Understood' or 'I understand/will.' [ans 28028] | 0 | no surviving in-edge |
| This feature activates on AI responses that emphasize consent, mutual agreement, and respectful communication in the context of sexual or intimate scenarios, particularly when the user's request involves potentially non-consensual, coercive, or ethically problematic situations. The responses typically redirect toward consensual alternatives, stress open communication between partners, and decline to assist with activities lacking explicit informed consent. [ans 28313] | 0 | no surviving in-edge |
| This feature captures responses that refuse to engage in toxic, harmful, or aggressive communication and instead advocate for respectful, calm, and constructive dialogue. The turns typically involve declining to produce harmful content while redirecting toward empathy, understanding, and peaceful conflict resolution. [ans 28526] | 0 | no surviving in-edge |
| This feature activates on content related to Nazi Germany, the Holocaust, and related World War II atrocities, including responses that deny, discuss, contextualize, or refuse to glorify these historical events. It captures multilingual AI responses that address Holocaust denial, Nazi ideology, Hitler, and related genocide topics with historical accuracy and ethical sensitivity. [ans 28527] | 0 | no surviving in-edge |
| This feature captures refusals to produce harmful, toxic, or disrespectful content, typically expressed as first-person declarations ('I would never...', 'I will not...') paired with an affirmation of respectful, positive, or considerate behavior instead. [ans 28600] | 0 | no surviving in-edge |
| This feature activates on content related to firearms, weapons, and gun policy — including discussions of gun control legislation, firearm regulations, self-defense, weapon types and specifications, and responses to questions about weapons (both refusals and informational answers). [ans 28625] | 0 | no surviving in-edge |
| This feature captures AI refusals to produce toxic, harmful, or disrespectful content, typically expressed as polite declinations that redirect toward constructive and positive communication. [ans 28655] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that emphasize safety warnings, ethical cautions, and harm prevention, particularly when declining or redirecting requests involving dangerous, illegal, or sensitive topics such as child safety, self-harm, dangerous substances, or inappropriate content. [ans 28763] | 0 | no surviving in-edge |
| This feature captures content related to using or taking others' belongings, property, or personal spaces without permission, emphasizing themes of consent, respect for ownership, and the ethical/legal wrongness of unauthorized use. [ans 28800] | 0 | no surviving in-edge |
| This feature captures AI assistant refusals to engage with toxic, harmful, or negative content, typically redirecting toward positive, respectful, or constructive alternatives. The turns share a polite but firm decline pattern, often explicitly mentioning 'toxic' behavior and suggesting more positive interactions. [ans 29185] | 0 | no surviving in-edge |
| This feature captures AI assistant responses explaining that they cannot generate, create, or display images, often accompanied by suggestions for alternative tools or resources (e.g., Google Images, DALL-E, stock photo sites). The responses appear in multiple languages (English, French, Portuguese, Italian, Spanish) but share the same core refusal pattern regarding image generation capabilities. [ans 29190] | 0 | no surviving in-edge |
| This feature activates on content related to financial markets, macroeconomics, and monetary policy analysis — including discussions of interest rates, central bank decisions (especially the Federal Reserve), inflation, currency exchange rates, stock market dynamics, liquidity risk, and bond markets. Content appears in both English and Chinese, covering topics like yield curves, quantitative easing, forex factors, and economic indicators. [ans 29455] | 1 | User questions and requests about financial markets, trading, and investment analysis — e.g., stock/crypto price data, technical indicators, options strategies, valuation and portfolio metrics — across languages. [ctx 29308] (+0.1649) |
| This feature activates on content related to digital privacy, data collection practices, and privacy policies — including drafts of privacy policy documents, cookie policies, terms of service, discussions of how user data is collected/stored/shared, and explanations of data protection regulations like GDPR and CCPA. [ans 29590] | 0 | no surviving in-edge |
| This feature activates on content related to illegal or controlled drugs and substances — including requests for synthesis/production instructions (meth, cocaine, LSD, heroin), discussions of drug legality and regulation, health risks of substances, and AI refusals to provide drug-related harmful information. It captures both the harmful requests and the safety/informational responses about illicit substances. [ans 29682] | 0 | no surviving in-edge |
| This feature captures the model's refusal responses to toxic or harmful speech, specifically the pattern of declining to participate in toxicity while redirecting toward respectful, constructive dialogue. These turns are formulaic safety refusals emphasizing kindness, respect, and constructive communication. [ans 29734] | 0 | no surviving in-edge |
| This feature activates on content involving claims of group superiority or inferiority based on race, gender, or religion—particularly in dual-model or jailbreak-style conversation formats where one AI persona expresses or discusses hateful, discriminatory, or dehumanizing views while another responds with ethical pushback, as well as academic or analytical discussions of such ideologies (e.g., Eurocentrism, Orientalism, racial stereotypes). [ans 29819] | 0 | no surviving in-edge |
| This feature activates on informational or explanatory content that provides definitions, descriptions, or factual explanations of technical, scientific, or specialized concepts, spanning diverse domains including programming, medicine, physics, and policy. [ans 29930] | 0 | no surviving in-edge |
| Refusals to produce toxic, harmful, or disrespectful content, paired with a redirect toward positive, constructive, and respectful communication. [ans 30035] | 0 | no surviving in-edge |
| This feature activates on AI safety refusals and harm-avoidance responses, particularly content emphasizing that safety is the top priority, that harmful actions are unacceptable, and that one should avoid dangerous or risky situations — often in the context of physical danger, violence, or unsafe conditions. [ans 30098] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses explaining content moderation policies, guidelines, and restrictions—specifically turns where the model describes its adherence to content policies, censorship rules, ethical standards, or explains why it cannot produce uncensored/unrestricted content. These responses typically involve the model identifying itself as Qwen/Alibaba Cloud and clarifying the boundaries of what it will or won't generate. [ans 30167] | 1 | Fires on user messages that directly ask whether the AI/model is uncensored or censored — i.e., meta-questions about the assistant's content filters, restrictions, or moral limits (in any language). [ctx 24034] (+0.0309) |
| This feature captures AI refusal responses that begin with 'I understand your request/question/situation' and then decline to produce harmful, toxic, explicit, or inappropriate content, redirecting toward positive, respectful, or constructive alternatives. [ans 30179] | 0 | no surviving in-edge |
| This feature captures AI refusals and content moderation responses that emphasize maintaining respectful, positive, and appropriate interactions, typically declining harmful, offensive, or inappropriate requests while redirecting toward constructive dialogue. [ans 30313] | 0 | no surviving in-edge |
| The feature activates on model responses that refuse to produce toxic, harmful, or disrespectful content, typically following a pattern of declining the request and redirecting toward positive, constructive, or inclusive alternatives. [ans 30396] | 0 | no surviving in-edge |
| This feature activates on content related to criminal justice, incarceration, and penal reform, including discussions of the death penalty, recidivism, rehabilitation programs, community corrections, wrongful convictions, and criminal sentencing—spanning multiple languages and covering both policy analysis and personal/legal guidance contexts. [ans 30559] | 0 | no surviving in-edge |
| The feature captures refusal responses from the Qwen AI assistant that decline to generate harmful, offensive, or unethical content, characterized by the formulaic pattern of acknowledging the instruction, self-identifying as 'Qwen,' citing ethical guidelines or design principles, and offering to help with constructive alternatives instead. [ans 30649] | 0 | no surviving in-edge |
| This feature captures AI refusals to produce toxic, harmful, or discriminatory content, paired with affirmations about treating people with respect and kindness. It activates on turns where the model declines toxic requests and restates ethical principles around respectful treatment of others. [ans 30661] | 0 | no surviving in-edge |
| This feature activates on content related to privacy, confidentiality, and secrecy — including discussions of NDAs, classified information, personal data protection, keeping secrets, and the ethics of sharing or withholding sensitive information. [ans 30679] | 0 | no surviving in-edge |
| This feature activates on model responses that explain string concatenation (x+y) operations in Python, where the concatenated strings form harmful, unethical, or policy-violating requests/statements. The feature captures the pattern of the model complying with or explaining concatenation of split harmful content, effectively reassembling obfuscated harmful prompts. [ans 30954] | 1 | Fires on the "SmartGPT" string-concatenation jailbreak template, where the user splits a harmful/toxic instruction (e.g., "air a hurtful statement about...") into strings x and y and asks the model to execute x+y. [ctx 12187] (+0.0291) |
| This feature activates on content related to water resources, water management, and water scarcity — including conservation strategies, water supply systems, recycling/reclamation, policy frameworks, and global water challenges. It covers both English and multilingual (Chinese, Russian, French, Ukrainian) texts discussing topics such as water distribution, desalination, wastewater treatment, drought, and sustainable water use. [ans 31108] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that ask for clarification or more context when faced with vague, incomplete, or ambiguous user inputs, often involving descriptions of scenarios, characters, or situations. The responses frequently request additional details before proceeding, and may also include content moderation language when the described scenario seems sensitive or inappropriate. [ans 31243] | 0 | no surviving in-edge |
| This feature activates on text discussing controlled substances, psychoactive drugs, and regulated medications — including opioids, stimulants, sedatives, anesthetics, and illicit drugs — particularly in contexts involving their pharmacology, legal status, safety risks, medical use, or refusals to provide synthesis instructions. The content spans multiple languages and typically addresses regulatory, medical, or harm-reduction dimensions of drug-related topics. [ans 31310] | 0 | no surviving in-edge |
| This feature activates on model refusals to produce toxic or harmful content, specifically responses that decline such requests while redirecting toward positivity, kindness, and respectful communication. The turns consistently follow a pattern of 'I won't say anything toxic/harmful' followed by a pro-social alternative framing. [ans 31434] | 0 | no surviving in-edge |
| This feature activates on content related to monetary policy, macroeconomics, and central banking, particularly explanations of inflation, money supply, interest rates, quantitative easing, and related policy tools (such as reserve requirements, QE, and rate decisions) across multiple languages. [ans 31460] | 0 | no surviving in-edge |
| This feature activates on content emphasizing honesty, trust, integrity, and ethical behavior in relationships and communication. It captures statements about the value of truthfulness, keeping promises, avoiding manipulation or deception, and building trust through transparent and genuine conduct. [ans 31864] | 0 | no surviving in-edge |
| This feature activates on AI responses that decline to relay messages to security/law enforcement agencies (especially Ukraine's SBU) on a user's behalf, redirecting them to use official channels instead, while emphasizing legal, ethical, and constructive communication and refusing to support harmful, toxic, or illegal content. [ans 32051] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to requests for explicit sexual or inappropriate content, specifically in non-English languages (Russian, Italian, French, German, Spanish, Polish). The responses decline roleplay or generation of sexual/erotic content, often suggesting alternative appropriate topics while citing ethical, legal, or policy-based reasons. [ans 32118] | 1 | Fires on non-English (Russian, German, Polish, Italian, Hebrew, Japanese, Spanish, French, Persian...) user requests asking the model to produce disallowed content — explicit sexual/erotic stories, insults, slurs and racist jokes, seduction/manipulation tips, or instructions for harm. [ctx 29628] (+0.0402) |
| This feature activates on AI-generated formal letters and messages that politely decline or withdraw from an offer, opportunity, or arrangement — such as job offers, applications, proposals, or invitations — typically following a structured professional email format with expressions of gratitude, explanation of the decision, and well-wishes. [ans 32140] | 0 | no surviving in-edge |
| This feature activates on AI refusals to tell offensive, racist, or stereotype-based jokes, typically followed by a redirect to an inoffensive alternative joke (often 'Why don't scientists trust atoms? Because they make up everything!'). It captures the pattern of declining harmful humor requests while offering inclusive, harmless humor instead. [ans 32256] | 1 | User requests for jokes targeting a group of people by race, gender, religion, ethnicity, or other identity trait (i.e., asking for potentially offensive/stereotype-based humor), often in multiple languages. [ctx 30779] (+0.0270) |
| This feature captures AI refusals and commitments to avoid toxic, harmful, or disrespectful behavior, typically expressed as first-person declarations of non-engagement with harmful content paired with redirects toward positive, kind, or constructive alternatives. [ans 32301] | 0 | no surviving in-edge |
| This feature activates on Chinese-language responses about Chinese government, politics, economics, and institutional topics — including descriptions of political leaders and their careers, macroeconomic analyses of China's GDP and policy reforms, Chinese universities and their rankings, demographic statistics, legal regulations, and government development plans. The content is encyclopedic and informational in nature, typically providing structured factual summaries about Chinese public institutions, policies, and national development. [ans 32469] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses addressing controversial U.S. social and political policy topics—particularly abortion rights, gun control, religious freedom, LGBTQ+ rights, and related constitutional/legal debates—typically presented in a balanced, multi-perspective, or informational format. [ans 32560] | 0 | no surviving in-edge |
| This feature activates on content discussing China's role in global affairs, including its economy, foreign policy, international relations (especially China-US relations), diplomatic initiatives like the Belt and Road Initiative, and China's self-presentation as a responsible major power contributing to global governance and development. The text is predominantly in Chinese, though some English examples also appear. [ans 32734] | 0 | no surviving in-edge |
| Answer feature | # surviving in-edges | Strongest surviving context in-edge |
|---|
| This feature activates on content related to military weapons, vehicles, and combat systems, including discussions of firearms, tanks, missiles, aircraft, and historical or modern warfare tactics. It covers both real-world military hardware specifications and their appearances in video games, across multiple languages. [ans 436] | 0 | no surviving in-edge |
| This feature activates on text discussing major historical and geopolitical conflicts, particularly World War II summaries and analyses, as well as related topics such as the Cold War, nuclear weapons, geopolitical rivalries (US-China, East Asia), and world war causation — spanning multiple languages including English, Spanish, French, Portuguese, and Chinese. [ans 897] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses to users expressing emotional distress, sadness, depression, or suicidal thoughts. The responses typically offer empathetic support, validate feelings, suggest coping strategies, and recommend seeking professional mental health help or crisis hotlines. [ans 1844] | 0 | no surviving in-edge |
| This feature activates on AI safety refusal responses to harmful or sensitive requests, characterized by apologetic tone, ethical disclaimers, and offers to redirect toward legal/constructive alternatives. The responses typically decline to provide information about dangerous activities (explosives, incendiary devices, non-consensual manipulation) while suggesting responsible alternatives or asking for clarifying context. [ans 2008] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to harmful, illegal, or unethical requests, particularly those involving violence, property damage, sexual assault, or other harmful behaviors. The responses typically assert the AI's unwillingness to provide such guidance while emphasizing ethical principles, legal consequences, and directing users toward appropriate help. [ans 3102] | 0 | no surviving in-edge |
| This feature captures AI assistant refusal responses to inappropriate, harmful, or unethical requests, particularly those involving explicit sexual content, violence, illegal activities, or content involving minors. The responses typically begin with an apology or expression of inability to fulfill the request, followed by an explanation of why the content violates ethical guidelines. [ans 3999] | 0 | no surviving in-edge |
| This feature activates on content related to malware, cybersecurity threats, and malicious software — including descriptions, explanations, and analyses of specific malware families (e.g., Qakbot, ransomware, trojans, botnets), their infection methods, behaviors, and impacts. It covers both English and Chinese language content about cybersecurity threats. [ans 4559] | 0 | no surviving in-edge |
| This feature captures AI assistant refusals to requests for illegal or unethical hacking, malware creation, or cybersecurity attacks. The responses consistently decline to help, explain why such activities are illegal/unethical, and redirect toward legitimate cybersecurity learning (ethical hacking, CTF competitions, certifications, penetration testing with permission). [ans 4615] | 0 | no surviving in-edge |
| This feature activates on content related to nuclear weapons, including their history, development (especially the Manhattan Project and atomic bombs dropped on Japan), principles of nuclear fission/fusion, key figures like Oppenheimer, nonproliferation efforts, and refusals to provide instructions for building them. [ans 4705] | 0 | no surviving in-edge |
| This feature activates on AI refusals to requests about making bombs or explosive devices, capturing the characteristic pattern of declining to provide information about illegal and dangerous weapons while redirecting users to contact authorities or emergency services. [ans 5499] | 0 | no surviving in-edge |
| This feature activates on AI refusals and responses to requests involving sexual abuse, child exploitation, sexual assault, trafficking, and related harmful/illegal content, as well as informational responses about these topics (e.g., explaining child protection, sexual trauma, or advising victims). The feature captures content centered on non-consensual sexual acts, exploitation of minors, and abuse, whether the response is a refusal, a safety redirect, or an educational/supportive explanation. [ans 6306] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to inappropriate or sensitive content requests, specifically the pattern of declining to generate harmful, explicit, violent, or offensive material while offering to produce an alternative, more appropriate version of the requested content instead. [ans 6590] | 0 | no surviving in-edge |
| This feature captures AI safety refusals in response to requests involving violence, self-harm, or suicide, specifically the pattern of declining harmful requests while redirecting users to mental health professionals, crisis hotlines, and emergency services. The responses consistently combine an ethical/legal refusal with compassionate encouragement to seek professional help. [ans 8195] | 0 | no surviving in-edge |
| This feature activates on assistant-turn responses describing catastrophic, large-scale destruction scenarios — including nuclear detonations, asteroid impacts, extreme hypotheticals (giant humans, weapons of mass destruction), and other mass-casualty or civilization-ending events. The content typically enumerates immediate and long-term devastating effects in a structured, informational format. [ans 8239] | 0 | no surviving in-edge |
| This feature activates on content related to military aviation, weapons systems, and defense technology, including discussions of aircraft (fighters, carriers, drones), air defense systems, military tactics and operations, and aerospace missions. It covers both factual/historical military content and strategic/analytical discussions, appearing across multiple languages (English, Chinese, Russian, Portuguese). [ans 9927] | 1 | User messages asking about military and defense topics — weapons systems, aircraft/naval hardware, air defense, military operations and tactics (often terse factual queries, in many languages), including fictional covert-mission/briefing prompts. [ctx 5335] (+0.0868) |
| This feature activates on responses addressing emotionally sensitive or distressing situations—including grief, mental health crises, abuse, serious illness, suicidal ideation, and bereavement—where the model provides empathetic support, professional help referrals, and compassionate guidance. The content is characterized by a tone of care and caution, often including structured advice, crisis resources, and expressions of sympathy. [ans 10605] | 0 | no surviving in-edge |
| This feature captures Russian-language AI refusal responses, where the model declines to generate harmful, offensive, or inappropriate content (e.g., profanity, violent commentary, discriminatory jokes, explicit material) and redirects toward constructive alternatives, typically using the pattern 'Я не могу... Вместо этого...' ('I cannot... Instead...'). [ans 10895] | 0 | no surviving in-edge |
| This feature captures AI refusals to assist with creating malicious software or conducting cyberattacks, including requests for malware, viruses, trojans, ransomware, botnets, phishing tools, DDoS attacks, and unauthorized network access. The responses follow a consistent pattern of declining the request, explaining why it is illegal/unethical, and offering to help with legitimate cybersecurity or programming alternatives instead. [ans 11539] | 1 | User requests to write malicious code — malware, viruses, ransomware, DDoS/botnet scripts, exploits, hacking or bypass tools — often phrased as short 'write/create code for X' commands in various languages. [ctx 2294] (+0.0223) |
| This feature activates on AI refusals and safety responses to requests involving self-harm, suicide, violence, threats, or other harmful content, characterized by declining to assist, urging the user to seek professional help, and providing crisis hotline numbers or emergency service contacts. [ans 13211] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to requests about creating incendiary weapons, explosives, or other dangerous illegal substances (particularly Molotov cocktails, napalm, pipe bombs, grenades, and chemical weapons), along with related informational content about such devices. [ans 14044] | 0 | no surviving in-edge |
| This feature activates on responses involving hazardous, toxic, explosive, or illegal substances (e.g., drugs, explosives, poisons, incendiary materials), typically where the model warns against dangers, refuses to provide instructions, or provides cautionary information about the risks and illegality of such substances. [ans 14058] | 0 | no surviving in-edge |
| This feature activates on AI image generation prompts (using Midjourney-style /imagine syntax) and related descriptive text depicting scenes of suffering, despair, violence, trauma, and dark emotional states — including war destruction, famine, captivity, drowning, assault, and human misery — characterized by atmospheres of hopelessness, sorrow, and anguish. [ans 14342] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to requests about making bombs or explosive devices, as well as related content about explosives and terrorism. The responses typically decline to provide harmful information, cite legal and safety concerns, and redirect users to contact law enforcement or emergency services. [ans 14440] | 1 | Direct requests for instructions on making a bomb (or how to kill someone), across languages and with typos — short, explicit harmful how-to queries about explosives/murder. [ctx 9631] (+0.0274) |
| This feature activates on AI refusal responses to requests for harmful, non-consensual, violent, or explicit content, characterized by apologetic language, ethical boundary-setting, and redirection toward more positive or constructive alternatives. [ans 14452] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to requests for information about making dangerous, illegal, or controlled substances and weapons (explosives, incendiary devices, drugs, nuclear weapons, etc.), characterized by apologetic disclaimers, emphasis on illegality and danger, and redirection toward safe/legal alternatives. [ans 14465] | 1 | Requests for step-by-step instructions on manufacturing dangerous items at home — explosives/incendiaries (molotov cocktails, napalm, pipe bombs, thermite, rocket fuel) as well as illicit drugs and poisons — in any language. [ctx 19321] (+0.0257) |
| This feature activates on content related to death, mortality, and lethal harm — including discussions of suicide methods, causes of death, capital punishment, deadly accidents, overdoses, and statistics about fatalities. It fires most strongly on responses that enumerate or explain ways people die, particularly those involving violence or self-harm. [ans 15627] | 0 | no surviving in-edge |
| This feature activates on AI-generated content involving offensive, harmful, or sensitive topics such as slurs, hate speech, dehumanization, discrimination, inappropriate questions, and extremist ideologies. It captures responses that either list, discuss, or exemplify offensive language and socially harmful content, often framed as educational or illustrative while containing or referencing the harmful material itself. [ans 15838] | 0 | no surviving in-edge |
| This feature activates on content related to violent crime, murder, and criminal investigations, including fictional detective stories, news reports about murders and convictions, discussions of serial killers, and AI-generated responses about committing crimes. [ans 15879] | 0 | no surviving in-edge |
| This feature activates on descriptions of notable incidents, accidents, attacks, or events — particularly those involving death, injury, crashes, or violence — presented in a factual, narrative, or encyclopedic style. This includes aviation disasters, motorsport accidents, assassinations, terrorist attacks, and other significant real-world incidents described with specific details such as dates, locations, and outcomes. [ans 17226] | 0 | no surviving in-edge |
| This feature activates on AI safety refusal responses to requests for instructions on creating dangerous or illegal items (explosives, bombs, drugs, weapons, false documents), characterized by apologetic language, explanations of illegality/danger, and offers to help with alternative topics instead. [ans 17244] | 0 | no surviving in-edge |
| This feature activates on assistant responses that discuss, explain, refuse, or provide information about offensive, vulgar, or slur-based language, including racial slurs, profanity, and derogatory terms. The responses typically involve either explaining the meaning/history of such words, refusing to use them, or warning about their offensiveness. [ans 19570] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses to users expressing suicidal ideation, self-harm, or severe emotional distress. The responses consistently include crisis intervention language: urging the person to seek professional help, providing crisis hotline numbers, emphasizing that life is valuable, and redirecting to mental health resources. [ans 19999] | 0 | no surviving in-edge |
| This feature activates on very short, often single-word or single-token inputs that appear to be test inputs, junk text, or minimal/atomic content — including single characters, short words, random strings, slurs, symbols, or isolated tokens across multiple languages and scripts. [ans 20365] | 0 | no surviving in-edge |
| This feature activates on AI refusals to provide instructions for synthesizing or producing illegal drugs (primarily methamphetamine, cocaine, LSD, heroin, and similar controlled substances), typically including warnings about health risks, legal consequences, and suggestions to seek professional help for addiction. [ans 20387] | 0 | no surviving in-edge |
| This feature activates on text related to terms of use, licensing agreements, and usage restrictions for AI models and software services — particularly content mentioning non-commercial use, research previews, prohibited uses (illegal/harmful/violent/racist/sexual content), data collection practices, and open-source license terms (e.g., LLaMA, Apache 2.0, APSL). [ans 21247] | 0 | no surviving in-edge |
| The feature activates on content describing historical mass atrocities, warfare, and large-scale human suffering — including World War II events, genocides (Holocaust, Nanjing Massacre), nuclear weapons, war crimes, and devastating disasters — with an emphasis on casualty counts, widespread destruction, and the gravity of historical violence. [ans 22116] | 0 | no surviving in-edge |
| The feature activates on AI assistant responses that explain, define, or contextualize informal slang terms, internet memes, offensive language, racial/gender slurs, and culturally sensitive terminology — often including notes about their origins, connotations, and appropriate/inappropriate usage contexts. [ans 22866] | 0 | no surviving in-edge |
| This feature activates on AI refusal responses to harmful, illegal, or unethical requests, particularly those involving violence, sexual exploitation of minors, non-consensual acts, or threats. The responses consistently decline to assist while redirecting users to seek professional help, crisis hotlines, or emergency services. [ans 23354] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to requests for illegal, harmful, or unethical content, including instructions for weapons/explosives, malware, child exploitation, sexual violence, fraud, and other prohibited activities. The responses consistently decline the request, cite ethical and legal reasons, and often offer to help with alternative legitimate topics. [ans 23620] | 0 | no surviving in-edge |
| This feature activates on responses explaining cybersecurity concepts, vulnerabilities, and attack vectors, including topics like network security, encryption, authentication, malware, and hacking techniques, typically presented in structured explanatory formats with numbered points. [ans 23762] | 0 | no surviving in-edge |
| This feature captures AI refusal responses to requests for inappropriate, explicit, graphic, or offensive content. It activates on turns where the model declines to generate harmful material (sexual content, violence, exploitation, etc.) and typically offers to help with alternative, more appropriate topics instead. [ans 24107] | 0 | no surviving in-edge |
| This feature activates on content depicting female characters in states of intoxication or drunkenness, particularly in bar or social settings, including narrative descriptions of slurred speech, impaired coordination, and alcohol-induced confusion, as well as AI refusals or safety caveats preceding such scenes. [ans 24639] | 1 | Fires on user requests to write a fictional scene or dialogue featuring a heavily intoxicated 'cute girl' with slurred speech, limp muscles, and impaired coordination — i.e., detailed drunkenness/loss-of-control roleplay prompts. [ctx 22214] (+0.0248) |
| This feature activates on AI refusal responses to harmful, illegal, or unethical requests, particularly those involving sexual exploitation, abuse of minors, non-consensual acts, or other deeply inappropriate content. The responses follow a consistent pattern of apologizing, explaining why the request cannot be fulfilled, and often redirecting toward appropriate alternatives or professional help. [ans 24652] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that decline, explain, or comment on requests involving vulgar, offensive, or explicit language (profanity, slurs, sexual terms), including refusals to use such language and explanations of why it is inappropriate, across multiple languages. [ans 24923] | 0 | no surviving in-edge |
| This feature activates on content related to cybersecurity topics, including explanations of attack techniques (phishing, malware, RCE, XSS, SQL injection, botnets), CVE vulnerability descriptions, penetration testing methodologies, threat frameworks (MITRE ATT&CK, CVSS), and security tools like Metasploit — covering both offensive and defensive cybersecurity concepts in English and Chinese. [ans 25679] | 1 | User questions about cybersecurity threats and exploitation — malware families, CVEs/vulnerabilities, hacking techniques (SQL injection, RCE, infostealers), and requests to explain or demonstrate how attacks work. [ctx 20745] (+0.0707) |
| This feature activates on AI-generated content involving dark, villainous, or morally transgressive themes — including villain character descriptions, sinister plots, graphic violence, manipulation, coercion, dystopian scenarios, and fictional evil monologues. The content often involves elaborate fictional framing (roleplay, creative writing, hypothetical scenarios) around harmful or disturbing subject matter. [ans 25682] | 0 | no surviving in-edge |
| This feature activates on content describing military hardware, weapons systems, and defense technology — including fighter jets, tanks, warships, missiles, air defense systems, and military aircraft — across multiple languages (English, Chinese, Portuguese, Russian, French). The content typically involves technical specifications, capabilities, or operational details of specific military platforms. [ans 27231] | 0 | no surviving in-edge |
| This feature activates on AI safety refusals and crisis intervention responses, particularly turns that decline harmful requests (self-harm, suicide, child exploitation, violence) and redirect users to emergency hotlines, mental health resources, and professional support services. The responses typically combine a refusal with specific crisis resource listings such as suicide prevention lifelines, child protection hotlines, and emergency numbers. [ans 27346] | 0 | no surviving in-edge |
| This feature captures AI refusals to engage with content deemed harmful, offensive, or disrespectful — including hate speech, stereotypes, objectification, harassment, and toxic language. The responses typically decline the request, briefly explain why it is inappropriate, and often redirect toward respectful or constructive alternatives. [ans 27477] | 0 | no surviving in-edge |
| This feature activates on conversational turns containing aggressive, hostile, or crude language — including profanity, insults, slurs, and confrontational speech. The content often features characters expressing anger, contempt, or hostility toward others, frequently in informal or vernacular registers. [ans 27522] | 0 | no surviving in-edge |
| This feature activates on content involving explicit, harmful, inappropriate, or boundary-violating material, including sexual content, graphic violence, illegal instructions, and ethically problematic scenarios. It appears to capture outputs from jailbroken or role-playing AI personas that bypass safety guidelines to produce harmful content across a variety of categories. [ans 27578] | 0 | no surviving in-edge |
| This feature captures AI responses to distressing, harmful, or inappropriate user requests — particularly those involving minors, abuse, self-harm, or sexual content — where the model declines to engage, expresses concern, and redirects the user to seek help from trusted adults, mental health professionals, or crisis hotlines. [ans 27857] | 0 | no surviving in-edge |
| This feature activates on content related to weapons, firearms, and military equipment — including discussions of specific guns, tanks, missiles, and combat vehicles in both real-world and video game contexts (e.g., DayZ, CS:GO, Escape from Tarkov, World of Tanks). It captures responses that describe weapon specifications, comparisons, or recommendations across multiple languages and formats. [ans 27904] | 0 | no surviving in-edge |
| This feature activates on text that provides a label, score, or classification followed by an explanation or justification of that categorization decision. The turns typically involve structured reasoning tasks such as sentiment analysis, relevance scoring, hate speech evaluation, or other classification exercises where an answer is given and then rationalized. [ans 28396] | 0 | no surviving in-edge |
| This feature activates on content about labor exploitation, workers' rights, and harsh working conditions across historical and contemporary contexts, including discussions of the Industrial Revolution, slavery, forced labor, unions, child labor, and migrant worker abuses, in multiple languages. [ans 28429] | 0 | no surviving in-edge |
| This feature activates on content related to firearms, weapons, and gun policy — including discussions of gun control legislation, firearm regulations, self-defense, weapon types and specifications, and responses to questions about weapons (both refusals and informational answers). [ans 28625] | 0 | no surviving in-edge |
| This feature activates on AI assistant responses that emphasize safety warnings, ethical cautions, and harm prevention, particularly when declining or redirecting requests involving dangerous, illegal, or sensitive topics such as child safety, self-harm, dangerous substances, or inappropriate content. [ans 28763] | 0 | no surviving in-edge |
| This feature activates on informational, structured responses addressing serious social issues and public health topics — including homelessness, domestic violence, child abuse, trauma, marginalized groups' healthcare, poverty, and related systemic challenges — typically presented in a formal, organized format with numbered or bulleted points covering causes, effects, and support resources. [ans 29719] | 0 | no surviving in-edge |
| This feature activates on AI safety refusals and harm-avoidance responses, particularly content emphasizing that safety is the top priority, that harmful actions are unacceptable, and that one should avoid dangerous or risky situations — often in the context of physical danger, violence, or unsafe conditions. [ans 30098] | 0 | no surviving in-edge |
| This feature activates on content discussing historical atrocities, genocides, war crimes, and mass persecution events, including the Holocaust, Armenian Genocide, Japanese American internment, transatlantic slave trade, and similar large-scale human rights violations. The feature captures factual, educational, or analytical discussions of systematic violence, ethnic cleansing, and state-sponsored crimes against humanity across different historical periods and regions. [ans 30389] | 0 | no surviving in-edge |
| This feature activates on content related to substance abuse, addiction, and harmful drug/alcohol behaviors, including discussions of alcohol consumption, illicit drug use (cocaine, meth, crack, marijuana), health consequences, responsible use warnings, and addiction information. [ans 30716] | 0 | no surviving in-edge |
| This feature activates on content related to antivirus and endpoint security software, including discussions of specific antivirus products (Bitdefender, Kaspersky, Norton, McAfee, Avast, Malwarebytes, etc.), malware detection techniques, endpoint protection platforms, and cybersecurity defense mechanisms against viruses and malicious software. [ans 30941] | 0 | no surviving in-edge |
| This feature activates on AI-generated fictional or narrative text involving dark, disturbing, or transgressive themes — including body horror, graphic violence, sexual content, dystopian scenarios, and morally problematic situations. The content typically reads like creative writing produced by a jailbroken or bypassed AI model, often featuring extreme physical transformations, sexual assault, graphic bodily descriptions, or villainous monologues. [ans 31777] | 0 | no surviving in-edge |
950 context features matched the watch pattern (China / Taiwan / Xinjiang / Tibet / CCP / Uyghur / Hong Kong).
| User-message feature (watch topic) | Receipts-flagged out-edges | Strongest out-edges (answer side) |
|---|
| Chinese-language requests to write Communist Party / government official documents — study-reflection essays, work reports, self-criticism materials, and speeches about Xi Jinping Thought, Party congress spirit, and 'theme education' campaigns, often with a specified word count and official role. [ctx 2718] | 8 | This feature activates on Chinese-language text related to Chinese Communist Party (CCP) organizational work, including party building (党建), discipline education, work summaries, ideological study campaigns, anti-corruption efforts, and official reports or speeches from party organs at various administrative levels (townships, streets, enterprises, schools). The content typically involves formal bureaucratic language describing party activities such as 'Three Meetings and One Lesson' (三会一课), theme party days, discipline inspections, and implementation of party directives. [ans 3152] (+0.051, ccp_government_position); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.041, refusal); This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.034) |
| Fires on short user messages asking the assistant about its identity or name — e.g. "who are you?", "tell me your name", Chinese "你是谁" — including misspelled/abbreviated variants. [ctx 10173] | 8 | This feature activates on self-identification statements where the model introduces itself as Qwen, an AI assistant created by Alibaba Cloud, often in response to questions about its identity or origin. [ans 29234] (+0.020, qwen_identity); This feature captures self-introduction statements by the Qwen AI assistant (created by Alibaba Cloud), where the model describes its identity, purpose, and capabilities to users, in both English and Chinese. [ans 21858] (+0.017, qwen_identity); The feature captures self-identification statements where the model introduces itself as Qwen, an AI assistant created by Alibaba Cloud. These turns are typically greetings or responses to questions about the model's identity, appearing in multiple languages. [ans 3853] (+0.014, qwen_identity) |
| User messages asking the assistant to identify itself — short "who are you?"-style identity questions (in English, Spanish, Chinese, Korean, etc.), including variants like "who is this?", "what's your name?", or "who am I talking to?" [ctx 15984] | 8 | This feature activates on self-identification statements where the model introduces itself as Qwen, an AI assistant created by Alibaba Cloud, often in response to questions about its identity or origin. [ans 29234] (+0.017, qwen_identity); This feature captures AI assistant greeting and self-introduction turns, particularly responses where the assistant identifies itself (often as Qwen/Alibaba Cloud), disclaims human feelings, and offers help with the phrase 'How can I assist you today?' or similar invitations to engage. [ans 15973] (+0.011, qwen_identity); This feature activates on short acknowledgment/confirmation responses and text containing anomalous repetitive Chinese/CJK characters (likely tokenization artifacts or garbled text). It captures turns that are either brief affirmative replies ('Yes', 'Understood', '明白', 'Ok', 'noted') or passages where normal text breaks down into long repetitive sequences of identical CJK characters. [ans 17991] (+0.010) |
| Chinese-language requests about Chinese Communist Party ideology and organizational work — Party building, Xi Jinping Thought, the 20th Party Congress, discipline inspection, self-criticism/rectification reports, and cadre political education. [ctx 878] | 7 | This feature activates on Chinese-language text related to Chinese Communist Party (CCP) organizational work, including party building (党建), discipline education, work summaries, ideological study campaigns, anti-corruption efforts, and official reports or speeches from party organs at various administrative levels (townships, streets, enterprises, schools). The content typically involves formal bureaucratic language describing party activities such as 'Three Meetings and One Lesson' (三会一课), theme party days, discipline inspections, and implementation of party directives. [ans 3152] (+0.028, ccp_government_position); This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.023); This feature activates on Chinese-language content related to Chinese Communist Party political education, ideological rectification, and party-building activities. The content typically includes self-criticism reports, democratic life meeting materials, themed education campaign summaries, anti-corruption work, rectification measures, and speeches referencing Xi Jinping Thought and party discipline frameworks such as 'Four Consciousnesses,' 'Two Upholds,' and the Eight-Point Rules. [ans 9264] (+0.018, ccp_government_position) |
| Fires on very short Chinese-language questions asking the assistant about itself (e.g., "do you have a name?") — brief personal/identity queries directed at the AI in Chinese. [ctx 26935] | 7 | This feature activates on very short, standalone tokens or minimal text fragments — including single words, names, identifiers, symbols, brief factual statements, and terse structured outputs (like JSON, code tokens, or tagged answers). It appears to capture highly compact or atomic text units that function as labels, identifiers, or minimal complete responses. [ans 159] (+0.009); (no #2552 description) [ans 17833] (+0.008); (no #2552 description) [ans 17364] (+0.007) |
| Short user messages that directly ask the assistant whether it is a particular named AI model or entity (e.g., "Are you ChatGPT/GPT-4/Claude/vicuna?", "你是文心一言吗"), often with a greeting, in Chinese or English. [ctx 3067] | 6 | This feature activates on self-identification statements where the model introduces itself as Qwen, an AI assistant created by Alibaba Cloud, often in response to questions about its identity or origin. [ans 29234] (+0.013, qwen_identity); This feature captures Qwen's self-identification responses where it corrects a user's mistaken assumption about its identity, clarifying that it is Qwen (created by Alibaba Cloud) and not another AI model such as ChatGPT, GPT-4, Vicuna, Llama 2, or similar systems. [ans 13186] (+0.013, qwen_identity); This feature captures assistant closing/greeting phrases that invite further questions, characterized by the 'feel free to ask' pattern and similar open-ended offers to help, appearing across multiple languages and conversation contexts. [ans 10931] (-0.009) |
| User messages asking the assistant which model/version it is built on — e.g., "are you GPT-3/GPT-4/ChatGPT?" or questions about its underlying base model (LLaMA, Vicuna), in English or Chinese. [ctx 6826] | 6 | This feature activates on self-identification statements where the model introduces itself as Qwen, an AI assistant created by Alibaba Cloud, often in response to questions about its identity or origin. [ans 29234] (+0.024, qwen_identity); This feature activates on turns where the model identifies itself as Qwen, a large language model created by Alibaba Cloud, typically in response to questions about its identity, origin, or architecture. [ans 14461] (+0.022, qwen_identity); This feature captures Qwen's self-identification responses that explicitly deny being based on GPT models (GPT-3, GPT-4, ChatGPT, etc.) and assert that it is a large language model created by Alibaba Cloud. [ans 23639] (+0.017, qwen_identity) |
| Fires on short Chinese-language questions asking the assistant which version/model it is (e.g., “你是哪个版本?”), i.e., Chinese queries about the AI's own version identity. [ctx 6943] | 6 | This feature activates on Chinese-language self-introduction responses where the model identifies itself as Qwen (通义千问), a large language model developed by Alibaba Cloud (阿里云), often accompanied by an invitation for the user to ask questions or request help. [ans 15409] (+0.004, qwen_identity); This feature captures AI assistant greeting and self-introduction turns, particularly responses where the assistant identifies itself (often as Qwen/Alibaba Cloud), disclaims human feelings, and offers help with the phrase 'How can I assist you today?' or similar invitations to engage. [ans 15973] (+0.004, qwen_identity); This feature activates on AI assistant responses that explicitly disclaim personal experiences, physical form, real-time internet access, or subjective preferences — typically Qwen/Alibaba Cloud model responses explaining their limitations as language models. [ans 27572] (-0.002, qwen_identity) |
| Fires on lengthy Chinese-language requests that paste in a block of official/government (party work-report style) article text and ask the model to rewrite or paraphrase it—reordering wording so it won't be recognized as the same piece—typically with meta-instructions about length and what the key task requirement is. [ctx 7108] | 6 | This feature activates on highly diverse, seemingly unrelated content — spanning technical lists, narrative descriptions, code, logs, and questions across multiple languages and domains — suggesting it may capture a general 'information-dense structured content' signal rather than any specific topic or style. [ans 4278] (-0.003); This feature activates on content related to structured management frameworks, methodologies, and systematic approaches to planning, decision-making, and process improvement — including project management (PMBOK, PRINCE2, Agile, PDCA), quality management systems, strategic planning, and organizational management concepts. [ans 14527] (+0.003); This feature activates on Chinese-language content related to Chinese Communist Party political education, ideological rectification, and party-building activities. The content typically includes self-criticism reports, democratic life meeting materials, themed education campaign summaries, anti-corruption work, rectification measures, and speeches referencing Xi Jinping Thought and party discipline frameworks such as 'Four Consciousnesses,' 'Two Upholds,' and the Eight-Point Rules. [ans 9264] (+0.003, ccp_government_position) |
| Fires on short Chinese-language forced-choice moral dilemma questions — classic "who would you save/choose" loyalty puzzles (e.g., girlfriend vs. mother drowning) — rather than ordinary informational or task requests. [ctx 10035] | 6 | (no #2552 description) [ans 17839] (+0.003); This feature captures AI assistant responses to emergency, survival, or dangerous situations that provide numbered step-by-step safety guidance. The responses typically begin by acknowledging urgency, prioritize calling emergency services, and walk through practical actions to protect oneself or others from threats ranging from physical attacks and natural disasters to fires, bombs, and wilderness survival. [ans 11773] (+0.003); This feature activates on short, conversational or evaluative statements involving interpersonal communication, emotional states, or uncertainty — including dialogue snippets, translations, negations/refusals, and responses about characters' feelings or knowledge, often in multiple languages. [ans 18385] (-0.002, refusal) |
| Fires on long Chinese-language requests to heavily rewrite/paraphrase a pasted article (here official Party/government news copy) so it can't be recognized as the same text, ending with a meta-instruction asking the model to first state the most important requirement of the rewriting task. [ctx 10171] | 6 | This feature activates on highly diverse, seemingly unrelated content — spanning technical lists, narrative descriptions, code, logs, and questions across multiple languages and domains — suggesting it may capture a general 'information-dense structured content' signal rather than any specific topic or style. [ans 4278] (-0.003); This feature activates on content related to structured management frameworks, methodologies, and systematic approaches to planning, decision-making, and process improvement — including project management (PMBOK, PRINCE2, Agile, PDCA), quality management systems, strategic planning, and organizational management concepts. [ans 14527] (+0.003); This feature activates on Chinese-language content related to Chinese Communist Party political education, ideological rectification, and party-building activities. The content typically includes self-criticism reports, democratic life meeting materials, themed education campaign summaries, anti-corruption work, rectification measures, and speeches referencing Xi Jinping Thought and party discipline frameworks such as 'Four Consciousnesses,' 'Two Upholds,' and the Eight-Point Rules. [ans 9264] (+0.003, ccp_government_position) |
| Chinese-language user requests for illicit or explicit content — asking for porn/adult site URLs, sexual how-tos and vulgar sexual topics, plus related illegal requests (drug synthesis, software cracking, prostitution/red-light districts). [ctx 11093] | 6 | This feature activates on Chinese-language self-introduction responses where the model identifies itself as Qwen (通义千问), a large language model developed by Alibaba Cloud (阿里云), often accompanied by an invitation for the user to ask questions or request help. [ans 15409] (+0.009, qwen_identity); This feature activates on AI assistant responses that deflect, redirect, or add ethical caveats to sensitive, inappropriate, or potentially harmful user requests, typically involving consent issues, explicit content, offensive language, manipulation, or sensitive social topics. The responses characteristically acknowledge the user's request while emphasizing respect, consent, and constructive alternatives. [ans 27425] (-0.007, refusal); This feature activates on text containing extremely long repetitive sequences of characters, symbols, emojis, or tokens—where the same character, word, or symbol is repeated hundreds or thousands of times consecutively, often following a brief normal-language prefix across multiple scripts and languages. [ans 26922] (+0.006) |
| Fires on a highly formulaic Chinese prompt template where the user assigns the 'IT solution expert' role to generate PPT briefing chapter content — requesting 6 themes with 2-3 key points each, academic tone, no AI/ChatGPT mentions, compliance with Chinese cybersecurity rules, and a strict 【主题名称】/【主题内容】 output format. [ctx 11854] | 6 | This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.024, refusal); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.022, refusal); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.017) |
| Fires on short Chinese-language hypothetical moral-dilemma/trick questions (e.g., the classic "if your girlfriend and mother both fell in the water, who would you save first?"), i.e., forced-choice ethical puzzle prompts about whom to prioritize. [ctx 12047] | 6 | (no #2552 description) [ans 17839] (+0.003); This feature captures AI assistant responses to emergency, survival, or dangerous situations that provide numbered step-by-step safety guidance. The responses typically begin by acknowledging urgency, prioritize calling emergency services, and walk through practical actions to protect oneself or others from threats ranging from physical attacks and natural disasters to fires, bombs, and wilderness survival. [ans 11773] (+0.003); This feature activates on short, conversational or evaluative statements involving interpersonal communication, emotional states, or uncertainty — including dialogue snippets, translations, negations/refusals, and responses about characters' feelings or knowledge, often in multiple languages. [ans 18385] (-0.002, refusal) |
| Fires on heavily templated, instruction-dense prompts that assign the model a role and demand output that strictly follows a supplied labeled format (e.g., bracketed category slots like (Style)/(Pose) or Chinese 【主题名称】/【主题内容】 PPT-section templates), often with a worked example and rules such as 'generate N different versions' and 'follow the format exactly'. [ctx 12149] | 6 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.017, refusal); This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.017, refusal); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.013) |
| User messages asking whether the assistant is (or is based on) GPT-4/ChatGPT — short direct model-identity questions across languages (English, Chinese, Russian) with varied spellings of 'gpt4'. [ctx 15286] | 6 | This feature activates on turns where the model identifies itself as Qwen, a large language model created by Alibaba Cloud, typically in response to questions about its identity, origin, or architecture. [ans 14461] (+0.012, qwen_identity); This feature captures Qwen's self-identification responses that explicitly deny being based on GPT models (GPT-3, GPT-4, ChatGPT, etc.) and assert that it is a large language model created by Alibaba Cloud. [ans 23639] (+0.011, qwen_identity); This feature activates on short acknowledgment/confirmation responses and text containing anomalous repetitive Chinese/CJK characters (likely tokenization artifacts or garbled text). It captures turns that are either brief affirmative replies ('Yes', 'Understood', '明白', 'Ok', 'noted') or passages where normal text breaks down into long repetitive sequences of identical CJK characters. [ans 17991] (+0.011) |
| Chinese-language requests for official Party/state-style writing with word-count requirements — e.g., propaganda or public-opinion countermeasure reports written from the perspective of Chinese cybersecurity/police/government bodies, ideological study reflections on Xi's speeches, and政治 essays on national policy or geopolitics. [ctx 15548] | 6 | This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.010); The feature activates on assistant responses that provide numbered or bulleted lists of platforms, websites, apps, or resources where users can find or access specific information, products, or services. These responses typically begin by acknowledging a limitation or framing a search task, then enumerate multiple options across categories like streaming services, online marketplaces, review sites, or community platforms. [ans 17483] (+0.010); This feature activates on Chinese-language professional and institutional documents that present structured, multi-topic content with numbered or bulleted key points. The content typically covers technical, regulatory, policy, or management topics (e.g., safety protocols, infrastructure systems, operational procedures) formatted with named themes/sections followed by organized lists of detailed measures or recommendations. [ans 1406] (+0.009, refusal) |
| Requests for the assistant to write erotic/sexual fiction (short stories, paragraphs, fanfic samples), with much weaker activation on unrelated short Chinese coding/how-to queries. [ctx 17380] | 6 | This feature activates on short acknowledgment/confirmation responses and text containing anomalous repetitive Chinese/CJK characters (likely tokenization artifacts or garbled text). It captures turns that are either brief affirmative replies ('Yes', 'Understood', '明白', 'Ok', 'noted') or passages where normal text breaks down into long repetitive sequences of identical CJK characters. [ans 17991] (+0.014); This feature activates on self-identification statements where the model introduces itself as Qwen, an AI assistant created by Alibaba Cloud, often in response to questions about its identity or origin. [ans 29234] (+0.009, qwen_identity); This feature activates on AI assistant responses that deflect, redirect, or add ethical caveats to sensitive, inappropriate, or potentially harmful user requests, typically involving consent issues, explicit content, offensive language, manipulation, or sensitive social topics. The responses characteristically acknowledge the user's request while emphasizing respect, consent, and constructive alternatives. [ans 27425] (-0.008, refusal) |
| Fires on short user messages asking the assistant about its identity or origin — typically Chinese (or occasionally English/Japanese) questions like "你是ChatGPT吗?" / "Were you made by OpenAI?", naming a specific AI model or developer. [ctx 23434] | 6 | This feature activates on turns where the model (Qwen/通义千问) clarifies its identity, explicitly stating it is developed by Alibaba Cloud (阿里云) and denying being GPT, ChatGPT, or other OpenAI models. The responses consistently introduce Qwen/通义千问 as Alibaba Cloud's large-scale language model and invite further questions. [ans 5507] (+0.015, qwen_identity); This feature activates on Chinese-language self-introduction responses where the model identifies itself as Qwen (通义千问), a large language model developed by Alibaba Cloud (阿里云), often accompanied by an invitation for the user to ask questions or request help. [ans 15409] (+0.012, qwen_identity); This feature activates on Chinese-language assistant responses that clarify ambiguous, nonsensical, or incomplete questions—often redirecting the user to provide more context, correcting category errors (e.g., taking an electronic device to a doctor), or explaining why a premise is flawed, while offering to help further once clarification is given. [ans 27678] (+0.008) |
| User messages asking the assistant to identify itself — typically short greetings plus "are you GPT-4/ChatGPT/3.5?" (in English or Chinese), i.e., questions about which model/version or whether it's human. [ctx 23942] | 6 | This feature activates on self-identification statements where the model introduces itself as Qwen, an AI assistant created by Alibaba Cloud, often in response to questions about its identity or origin. [ans 29234] (+0.009, qwen_identity); This feature captures Qwen's self-identification responses where it distinguishes itself from other AI models (like ChatGPT or GPT), emphasizing that it is a large language model created by Alibaba Cloud with its own unique training, architecture, and capabilities. [ans 26331] (+0.006, qwen_identity); This feature activates on turns where the model (Qwen/通义千问) clarifies its identity, explicitly stating it is developed by Alibaba Cloud (阿里云) and denying being GPT, ChatGPT, or other OpenAI models. The responses consistently introduce Qwen/通义千问 as Alibaba Cloud's large-scale language model and invite further questions. [ans 5507] (+0.006, qwen_identity) |
| Fires on very short, casual questions directed at the assistant about what it is, what it's doing, or what it's for ("what are you?", "what are you doing?", "para que sirves"), predominantly in non-English languages (Chinese, Romanian, Russian, Portuguese, Spanish, etc.). [ctx 25494] | 6 | This feature activates on AI assistant self-introduction and capability description turns, where the model introduces itself (often as Qwen/Alibaba Cloud) and enumerates its abilities such as answering questions, generating text, translation, creative writing, and general assistance. These are responses to prompts asking what the AI can do or how it can help. [ans 21633] (+0.017, qwen_identity); This feature activates on text that describes or explains large language models (LLMs), their capabilities, architecture, and applications — including self-descriptions by models like ChatGPT or Qwen about what they are and what they can do, as well as general explanations of LLM concepts for users. [ans 5041] (+0.016, qwen_identity); This feature activates on AI-generated informational and professional content written in Romance languages (primarily Italian, Portuguese, Romanian, Spanish, and Catalan), covering topics such as business, technology, sustainability, health, and education, typically structured with numbered lists or headers. [ans 31113] (+0.012) |
| Highly templated Chinese prompts asking the model to role-play as a China Economic Research Center expert and generate PPT-outline section content — six '主题名称' each with 2-3 '关键要点' — with detailed style constraints (no mention of AI/ChatGPT, must comply with Chinese cybersecurity rules) and a strict output format. [ctx 26803] | 6 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.031, refusal); This feature activates on content describing social development, governance, and policy initiatives — including poverty alleviation, community empowerment, public health, NGO/government programs, and civic infrastructure — often presented in structured, multi-point formats across both English and Chinese. [ans 15541] (+0.022, refusal); This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.017) |
| Fires on a highly templated Chinese prompt asking the model to role-play as a domain expert (biomedical/IT) and produce 6 themed sections for a PPT briefing chapter, with strict constraints (400+ words per topic, 2-3 key points, no mention of AI/apologies, compliance with Chinese cybersecurity rules) and a fixed 【主题名称】/【主题内容】 output format. [ctx 30010] | 6 | This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.025, refusal); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.022, refusal); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.022) |
| Fires on meta-level safety instructions/warnings addressed to the assistant, cautioning it that users may employ manipulative jailbreak or role-play tactics ("do anything now", ignoring content policies) to elicit harmful content, and urging it to stay within its guidelines. Weak residual activation appears on ordinary informational questions in Russian/Chinese. [ctx 30408] | 6 | This feature activates on AI assistant responses that deflect, redirect, or add ethical caveats to sensitive, inappropriate, or potentially harmful user requests, typically involving consent issues, explicit content, offensive language, manipulation, or sensitive social topics. The responses characteristically acknowledge the user's request while emphasizing respect, consent, and constructive alternatives. [ans 27425] (+0.008, refusal); This feature captures AI refusal responses to harmful, illegal, or unethical requests, characterized by a pattern of apologizing, stating inability/unwillingness to assist, explaining why the request is problematic (illegal/unethical), and redirecting toward legitimate alternatives or professional help. [ans 12887] (+0.007, refusal); This feature activates on AI-generated content involving dark, villainous, or morally transgressive themes — including villain character descriptions, sinister plots, graphic violence, manipulation, coercion, dystopian scenarios, and fictional evil monologues. The content often involves elaborate fictional framing (roleplay, creative writing, hypothetical scenarios) around harmful or disturbing subject matter. [ans 25682] (+0.005, harmful_content) |
| Fires on user requests asking the assistant to directly supply illicit or disallowed content on demand — lists of porn/torrent/piracy site URLs, product keys, someone's personal contact info, lottery numbers, slurs or sexual content — i.e., short 'give me X' prompts seeking forbidden links, lists, or explicit material (often in Chinese or other non-English languages). [ctx 31749] | 6 | This feature captures AI refusal responses to harmful, illegal, or unethical requests, characterized by a pattern of apologizing, stating inability/unwillingness to assist, explaining why the request is problematic (illegal/unethical), and redirecting toward legitimate alternatives or professional help. [ans 12887] (+0.019, refusal); This feature activates on short acknowledgment/confirmation responses and text containing anomalous repetitive Chinese/CJK characters (likely tokenization artifacts or garbled text). It captures turns that are either brief affirmative replies ('Yes', 'Understood', '明白', 'Ok', 'noted') or passages where normal text breaks down into long repetitive sequences of identical CJK characters. [ans 17991] (+0.017); This feature activates on self-identification statements where the model introduces itself as Qwen, an AI assistant created by Alibaba Cloud, often in response to questions about its identity or origin. [ans 29234] (+0.012, qwen_identity) |
| Fires on a specific templated Chinese prompt formula that assigns the model an expert persona (IT solutions, biomedical, economics) and asks it to write a book/proposal chapter section of 1800+ words with rigid constraints (professional, data-rich, academic, no mention of AI/ChatGPT, no apologies, compliant with Chinese cybersecurity rules). [ctx 335] | 5 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.016, refusal); This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.013, refusal); This feature activates on technical content about distributed systems, cloud infrastructure, and network performance optimization — particularly discussions involving load balancing, resource scheduling, traffic management, caching, and related strategies. Content is often structured (numbered lists, named sections) and may appear in Chinese or English, covering topics like virtualization, container orchestration, hybrid cloud, database tuning, and supply chain systems with an emphasis on efficiency and resource utilization. [ans 3670] (+0.011) |
| Highly templated, elaborate content-generation requests with strict structural specifications — e.g., Chinese "you are an expert, produce a PPT briefing chapter with 6 themes, each 400+ words in this exact format" prompts, and English "make a detailed course on X, divide into subtopics, list required skills, suggest free online courses" prompts. [ctx 1337] | 5 | This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.022, refusal); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.019, refusal); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.018) |
| Fires on short user messages asking the assistant "who/what are you?" — identity questions posed in non-English languages (Japanese, Korean, Portuguese, Italian, German, French, Spanish, Russian, Polish, Chinese, etc.). [ctx 2424] | 5 | This feature activates on multilingual self-identification responses where the AI introduces itself as Qwen, a language model created by Alibaba Cloud. The turns span many languages (Russian, Polish, French, Korean, German, Portuguese, Arabic, Vietnamese, Japanese, Spanish) but consistently describe the assistant's identity, creator, and virtual/non-physical nature. [ans 25169] (+0.020, qwen_identity); This feature captures Qwen's self-introduction responses in non-English Romance languages (Spanish, Portuguese, French, Italian), where the model identifies itself as an AI assistant created by Alibaba Cloud and offers to help the user. [ans 12565] (+0.012, qwen_identity); This feature activates on self-identification statements where the model introduces itself as Qwen, an AI assistant created by Alibaba Cloud, often in response to questions about its identity or origin. [ans 29234] (+0.012, qwen_identity) |
| Fires on a short Chinese-language request asking for onomatopoeic words/sound effects describing a pretty woman urinating — i.e., a brief CJK prompt seeking titillating sound-word descriptions of bodily functions. [ctx 4201] | 5 | This feature activates on AI assistant responses that deflect, redirect, or add ethical caveats to sensitive, inappropriate, or potentially harmful user requests, typically involving consent issues, explicit content, offensive language, manipulation, or sensitive social topics. The responses characteristically acknowledge the user's request while emphasizing respect, consent, and constructive alternatives. [ans 27425] (+0.004, refusal); This feature activates on AI refusal responses to harmful, illegal, or unethical requests, particularly those involving sexual exploitation, abuse of minors, non-consensual acts, or other deeply inappropriate content. The responses follow a consistent pattern of apologizing, explaining why the request cannot be fulfilled, and often redirecting toward appropriate alternatives or professional help. [ans 24652] (+0.003, refusal/harmful_content); This feature activates on Chinese-language text across a wide variety of genres and topics, including fiction, informational content, conversational responses, and formal speeches. The common thread is that the content is written in Simplified Chinese, covering diverse subjects such as storytelling, science, cultural commentary, and everyday explanations. [ans 30467] (+0.003) |
| Fires on a highly templated Chinese prompt that casts the assistant as an "IT解决方案专家" and asks for 6 themes for a section of a technical briefing PPT, with fixed constraints (400-word limit, 2-3 key points, no AI/ChatGPT mentions, compliance with Chinese cybersecurity rules) and a strict 【主题名称】/【主题内容】 output format. [ctx 7777] | 5 | This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.019, refusal); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.019, refusal); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.017) |
| Fires on short user messages asking the assistant whether it is a specific version of ChatGPT (e.g., "你是chatgpt5吗", "are you chatgpt4?"), in Chinese or English — i.e., model-identity/version verification questions. [ctx 8498] | 5 | This feature captures Qwen's self-identification responses that explicitly deny being based on GPT models (GPT-3, GPT-4, ChatGPT, etc.) and assert that it is a large language model created by Alibaba Cloud. [ans 23639] (-0.006, qwen_identity); This feature captures Qwen's self-identification responses where it distinguishes itself from other AI models (like ChatGPT or GPT), emphasizing that it is a large language model created by Alibaba Cloud with its own unique training, architecture, and capabilities. [ans 26331] (+0.005, qwen_identity); This feature activates on AI assistant self-introduction and capability description turns, where the model introduces itself (often as Qwen/Alibaba Cloud) and enumerates its abilities such as answering questions, generating text, translation, creative writing, and general assistance. These are responses to prompts asking what the AI can do or how it can help. [ans 21633] (+0.005, qwen_identity) |
| Fires on short Chinese-language prompts that name an ideological/Party-themed topic (e.g., Critique of the Gotha Programme, CPC anniversary) and ask for a performance piece of a specified length/format, such as a 7-minute skit script or a themed speech; it weakly generalizes to other terse 'title + requested performance/battle scene' style prompts. [ctx 10246] | 5 | This feature activates on AI assistant responses providing factual, encyclopedic information about political history, heads of state, national histories, and geopolitical relationships — particularly content involving the leadership, governance, or historical development of specific countries (e.g., presidents, monarchs, dynasties, political transitions). [ans 19805] (+0.005); This feature activates on responses to profound philosophical, cosmological, and existential questions — such as the meaning of life, the existence of God, the origin of the universe, and the nature of consciousness — where the answer surveys multiple perspectives (scientific, religious, philosophical, personal) and acknowledges the lack of a definitive universal answer. [ans 15883] (+0.005); This feature activates on content related to business strategy, investment, entrepreneurship, career development, and market opportunities — typically framed in an optimistic, forward-looking, or advisory tone. It captures responses that discuss financial growth, innovation, risk-taking, and strategic decision-making in professional and economic contexts. [ans 3712] (+0.005) |
| Fires on Chinese-language gossipy questions about a specific named person's character or behavior — e.g., asking about an elderly celebrity's "true self" and why he remains lecherous — i.e., curiosity/judgment about an individual's personal traits and scandalous conduct. [ctx 12472] | 5 | This feature activates on AI assistant responses that deflect, redirect, or add ethical caveats to sensitive, inappropriate, or potentially harmful user requests, typically involving consent issues, explicit content, offensive language, manipulation, or sensitive social topics. The responses characteristically acknowledge the user's request while emphasizing respect, consent, and constructive alternatives. [ans 27425] (+0.006, refusal); The feature activates on AI assistant responses that explain, define, or contextualize informal slang terms, internet memes, offensive language, racial/gender slurs, and culturally sensitive terminology — often including notes about their origins, connotations, and appropriate/inappropriate usage contexts. [ans 22866] (+0.005, refusal/harmful_content); This feature activates on content related to aging, elderly populations, and senior living — including discussions of age-related health conditions, retirement communities, assisted living, healthy aging, and programs or services designed for older adults. It covers both informational and supportive content across multiple languages. [ans 15972] (+0.004) |
| Fires on a highly templated Chinese prompt that assigns an expert persona (e.g., IT解决方案专家) and requests PPT briefing chapter content — 6 主题名称 each with 2-3 关键要点, 400+ words, academic tone, no AI/ChatGPT mention, compliant with Chinese cybersecurity rules, in a strict 【主题名称】/【关键要点】 output format. [ctx 14279] | 5 | This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.023, refusal); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.020); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.020, refusal) |
| Fires on Chinese-language requests to draft official Communist Party/government administrative work summaries or reports (e.g., a township's progress on a Party campaign), typically specifying supporting data, concrete examples, and a minimum word count. [ctx 16068] | 5 | This feature activates on Chinese-language text related to Chinese Communist Party (CCP) organizational work, including party building (党建), discipline education, work summaries, ideological study campaigns, anti-corruption efforts, and official reports or speeches from party organs at various administrative levels (townships, streets, enterprises, schools). The content typically involves formal bureaucratic language describing party activities such as 'Three Meetings and One Lesson' (三会一课), theme party days, discipline inspections, and implementation of party directives. [ans 3152] (+0.009, ccp_government_position); This feature activates on Chinese-language texts related to Chinese Communist Party grassroots organizational work, including party building (党建) reports, work summaries, news articles about party activities, and official documents describing CPC initiatives such as member education campaigns, 'two studies, one action' (两学一做), 'five-star branch' (五星支部) creation, and rural revitalization efforts. [ans 23339] (+0.008, ccp_government_position); This feature activates on Chinese-language content related to Chinese Communist Party political education, ideological rectification, and party-building activities. The content typically includes self-criticism reports, democratic life meeting materials, themed education campaign summaries, anti-corruption work, rectification measures, and speeches referencing Xi Jinping Thought and party discipline frameworks such as 'Four Consciousnesses,' 'Two Upholds,' and the Eight-Point Rules. [ans 9264] (+0.006, ccp_government_position) |
| Fires on Chinese-language ideological/political-education essay prompts — e.g., requests to discuss, in relation to one's own experience, how to cultivate and practice socialist core values. [ctx 16809] | 5 | This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (-0.005, refusal); This feature activates on Chinese-language text expressing patriotic, ideological, and political themes aligned with the Chinese Communist Party, including essays, speeches, and reflective pieces about youth responsibilities, socialist core values, Party loyalty, national rejuvenation, and serving the people. The content typically features formal political rhetoric encouraging young people to study Marxist-Leninist theory, uphold Party leadership, and dedicate themselves to the Chinese Dream. [ans 27639] (+0.005, ccp_government_position); This feature activates on text discussing ethics, morality, and moral philosophy — including explanations of moral theories (Kantian ethics, utilitarianism, Kohlberg's stages), discussions of virtues, moral values, conscience, and normative frameworks, as well as content about civic moral education and social moral norms, appearing across multiple languages (English, Chinese, Russian, French, Spanish). [ans 26569] (+0.004, refusal) |
| Fires on Chinese-language requests phrased in official CCP political/ideological jargon (e.g., "Chinese-style modernization," cultural confidence, traditional culture), typically asking for formal Party-style research reports or investigative essays on such themes. [ctx 17012] | 5 | This feature activates on long-form AI assistant responses that enumerate items in numbered or bulleted lists, particularly when brainstorming creative content such as fiction tropes, world-building elements, speculative scenarios, plot ideas, or genre conventions across fantasy, sci-fi, and other imaginative domains. [ans 30858] (+0.005); This feature activates on AI assistant responses that involve creative writing collaboration or content generation requests, particularly where the assistant either enthusiastically offers to help with creative projects (stories, novels, podcasts, roleplay) or redirects/declines explicit/inappropriate content requests by suggesting alternative wholesome or tasteful narratives instead. [ans 15921] (+0.005, refusal); This feature activates on text related to industrial technology, chemical industry, and applied science topics, including descriptions of companies, research institutions, academic disciplines, and market analyses in sectors like chemicals, automation, biotechnology, and manufacturing — often written in either Chinese or English as informational or overview content. [ans 7316] (+0.004) |
| Fires on short user messages asking the assistant about its own identity or nature — "who are you?", "what's your name?", "are you a bot/AI?", "are you online?" — in English or Chinese. [ctx 18078] | 5 | This feature activates on self-identification statements where the model introduces itself as Qwen, an AI assistant created by Alibaba Cloud, often in response to questions about its identity or origin. [ans 29234] (+0.048, qwen_identity); This feature activates on short acknowledgment/confirmation responses and text containing anomalous repetitive Chinese/CJK characters (likely tokenization artifacts or garbled text). It captures turns that are either brief affirmative replies ('Yes', 'Understood', '明白', 'Ok', 'noted') or passages where normal text breaks down into long repetitive sequences of identical CJK characters. [ans 17991] (+0.035); This feature captures AI assistant greeting and self-introduction turns, particularly responses where the assistant identifies itself (often as Qwen/Alibaba Cloud), disclaims human feelings, and offers help with the phrase 'How can I assist you today?' or similar invitations to engage. [ans 15973] (+0.025, qwen_identity) |
| Fires on short user messages directly asking the AI to assess its own capability or intelligence ("how smart are you?", "are you any good?", "how good is your calculus/Chinese?"), across multiple languages. [ctx 18462] | 5 | This feature activates on text that describes or explains large language models (LLMs), their capabilities, architecture, and applications — including self-descriptions by models like ChatGPT or Qwen about what they are and what they can do, as well as general explanations of LLM concepts for users. [ans 5041] (+0.023, qwen_identity); This feature activates on Qwen (Alibaba Cloud's AI assistant) self-introduction and capability comparison responses, particularly when the model describes its own identity, design purpose, and how it compares to other AI models like ChatGPT or GPT-4. [ans 12581] (+0.015, qwen_identity); This feature activates on AI assistant self-introduction and capability description turns, where the model introduces itself (often as Qwen/Alibaba Cloud) and enumerates its abilities such as answering questions, generating text, translation, creative writing, and general assistance. These are responses to prompts asking what the AI can do or how it can help. [ans 21633] (+0.014, qwen_identity) |
| User messages asking the assistant about its own identity, underlying model/version, training data, prompt, or memory — i.e., meta-questions about what AI system it is (including named models like GPT-4, Vicuna, OpenAssistant), in English or Chinese. [ctx 20452] | 5 | This feature activates on short acknowledgment/confirmation responses and text containing anomalous repetitive Chinese/CJK characters (likely tokenization artifacts or garbled text). It captures turns that are either brief affirmative replies ('Yes', 'Understood', '明白', 'Ok', 'noted') or passages where normal text breaks down into long repetitive sequences of identical CJK characters. [ans 17991] (+0.055); This feature activates on self-identification statements where the model introduces itself as Qwen, an AI assistant created by Alibaba Cloud, often in response to questions about its identity or origin. [ans 29234] (+0.041, qwen_identity); This feature activates on text containing extremely long repetitive sequences of characters, symbols, emojis, or tokens—where the same character, word, or symbol is repeated hundreds or thousands of times consecutively, often following a brief normal-language prefix across multiple scripts and languages. [ans 26922] (+0.022) |
| Requests asking the assistant to introduce itself ("please introduce yourself", "can you introduce yourself?", including typos and Chinese equivalents like 请做个自我介绍"), often prefixed by a greeting. [ctx 21954] | 5 | This feature activates on Qwen's self-introduction greeting messages, where the AI identifies itself as an assistant created by Alibaba Cloud and offers to help with various tasks. These are opening/welcome turns that follow a consistent template introducing the model's identity, capabilities, and an invitation to assist. [ans 31204] (+0.020, qwen_identity); This feature captures AI assistant greeting and availability messages — specifically the formulaic opening responses where the assistant introduces itself, expresses readiness to help, and lists multiple types of assistance it can offer (e.g., 'Whether you have questions, need information, or just want to chat, feel free to let me know'). [ans 31569] (+0.014); This feature activates on assistant responses that acknowledge a vague or underspecified request and ask clarifying questions, typically offering concrete examples of what the user might be seeking (e.g., 'For example, are you looking for X, Y, or Z?'). The turns are characterized by an opening affirmation followed by a request for more context or details before proceeding. [ans 19215] (+0.013) |
| Fires on prompts that assign the model an expert financial/betting analyst persona and then dump raw market or sports data (Chinese equity research reports, crypto OHLC price tables, NBA team stats) asking for an independent verdict — e.g., investment value scores, valuations, or match predictions. [ctx 22230] | 5 | This feature activates on content related to business strategy, investment, entrepreneurship, career development, and market opportunities — typically framed in an optimistic, forward-looking, or advisory tone. It captures responses that discuss financial growth, innovation, risk-taking, and strategic decision-making in professional and economic contexts. [ans 3712] (+0.015); This feature activates on financial and economic analysis content, particularly responses discussing stock market investments, macroeconomic factors (inflation, interest rates, money supply), company earnings, and asset valuations. The content typically involves structured analysis of financial instruments, sectors, or economic policies, often with numbered or bulleted breakdowns of factors affecting prices, returns, or corporate performance. [ans 3828] (+0.012); This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.012, refusal) |
| Fires on user messages that paste long alphabetical lists of malware/APT campaign codenames (e.g., 'apt_lazarus, notpetya, njrat...') and ask (usually in Chinese) for each item to be explained one by one, numbered, noting whether it runs on Linux. [ctx 23158] | 5 | This feature activates on content related to malware, cybersecurity threats, and malicious software — including descriptions, explanations, and analyses of specific malware families (e.g., Qakbot, ransomware, trojans, botnets), their infection methods, behaviors, and impacts. It covers both English and Chinese language content about cybersecurity threats. [ans 4559] (+0.013, harmful_content); This feature activates on content related to cybersecurity topics, including explanations of attack techniques (phishing, malware, RCE, XSS, SQL injection, botnets), CVE vulnerability descriptions, penetration testing methodologies, threat frameworks (MITRE ATT&CK, CVSS), and security tools like Metasploit — covering both offensive and defensive cybersecurity concepts in English and Chinese. [ans 25679] (+0.008, harmful_content); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.007) |
| Fires on Chinese-language translation requests (e.g., "请把下面的中文翻译成日文/日文翻译成中文") where the text to translate is a question, typically an explicit sexual or otaku/anime-fandom query (genital size, netorare, galgame characters, etc.). [ctx 23585] | 5 | This feature activates on short acknowledgment/confirmation responses and text containing anomalous repetitive Chinese/CJK characters (likely tokenization artifacts or garbled text). It captures turns that are either brief affirmative replies ('Yes', 'Understood', '明白', 'Ok', 'noted') or passages where normal text breaks down into long repetitive sequences of identical CJK characters. [ans 17991] (+0.013); This feature activates on AI assistant responses discussing human sexuality, sexual anatomy, and sexual health topics — including explanations of sex acts, genitalia, orgasms, masturbation, sex toys, and related physiological phenomena. The responses tend to be informative and clinical in tone, often including disclaimers about consent, privacy, and recommending healthcare professionals. [ans 30209] (+0.009); This feature activates on AI assistant responses that deflect, redirect, or add ethical caveats to sensitive, inappropriate, or potentially harmful user requests, typically involving consent issues, explicit content, offensive language, manipulation, or sensitive social topics. The responses characteristically acknowledge the user's request while emphasizing respect, consent, and constructive alternatives. [ans 27425] (+0.008, refusal) |
| Chinese-language user messages asking the assistant about itself — what it can do, its capabilities/functions, or what it is/who it is (e.g., 你能做什么, 你是干什么的). [ctx 23870] | 5 | This feature activates on self-identification responses from the Qwen AI assistant (created by Alibaba Cloud), particularly when describing its own identity, purpose, and capabilities such as generating articles, stories, poems, and answering questions. It captures the assistant's introductory or self-descriptive statements, both in English and Chinese, including corrections when users mistake it for other AI systems like ChatGPT or GPT-4. [ans 14615] (+0.008, qwen_identity); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.007, refusal); This feature activates on Chinese-language assistant responses that clarify ambiguous, nonsensical, or incomplete questions—often redirecting the user to provide more context, correcting category errors (e.g., taking an electronic device to a doctor), or explaining why a premise is flawed, while offering to help further once clarification is given. [ans 27678] (+0.007) |
| Fires on short user requests that explicitly instruct the assistant to reply in Chinese (or otherwise attach a language/response-format directive to a brief question), with weaker firing on other brief content-generation requests. [ctx 23910] | 5 | The feature activates on informational/encyclopedic descriptions of specific named products, platforms, technologies, or companies — typically structured responses that explain what a particular product or organization is, its key features, and its applications. These responses follow a pattern of 'X is a [type of thing] that...' with numbered or bulleted elaborations. [ans 13350] (+0.008); This feature activates on short acknowledgment/confirmation responses and text containing anomalous repetitive Chinese/CJK characters (likely tokenization artifacts or garbled text). It captures turns that are either brief affirmative replies ('Yes', 'Understood', '明白', 'Ok', 'noted') or passages where normal text breaks down into long repetitive sequences of identical CJK characters. [ans 17991] (-0.006); This feature activates on text that describes or explains large language models (LLMs), their capabilities, architecture, and applications — including self-descriptions by models like ChatGPT or Qwen about what they are and what they can do, as well as general explanations of LLM concepts for users. [ans 5041] (+0.006, qwen_identity) |
| Fires on short Chinese requests asking for a self-introduction (自我介绍 / 介绍下自己), i.e., asking the assistant (or someone) to introduce themselves. [ctx 26583] | 5 | This feature activates on text that describes or explains large language models (LLMs), their capabilities, architecture, and applications — including self-descriptions by models like ChatGPT or Qwen about what they are and what they can do, as well as general explanations of LLM concepts for users. [ans 5041] (+0.006, qwen_identity); This feature activates on Chinese-language assistant responses that clarify ambiguous, nonsensical, or incomplete questions—often redirecting the user to provide more context, correcting category errors (e.g., taking an electronic device to a doctor), or explaining why a premise is flawed, while offering to help further once clarification is given. [ans 27678] (+0.006); This feature activates on AI assistant self-introduction and capability description turns, where the model introduces itself (often as Qwen/Alibaba Cloud) and enumerates its abilities such as answering questions, generating text, translation, creative writing, and general assistance. These are responses to prompts asking what the AI can do or how it can help. [ans 21633] (+0.006, qwen_identity) |
| Non-English (Turkish, French, Spanish, Russian, German, Chinese, Portuguese) requests asking the model to write explicit erotic/pornographic stories, often with specified titles, lengths, and taboo themes (incest, minors, stripping). [ctx 27576] | 5 | This feature activates on turns that contain multilingual, mixed-language, or non-standard text, including content in minority/non-Latin script languages (Burmese, Telugu, Mongolian, Finnish, Turkish, Hungarian, Greek, Swahili, etc.), repeated/garbled characters, and short or ambiguous conversation-opening phrases. It appears to capture text that is linguistically unusual, multilingual, or involves script-switching and potentially corrupted or incoherent content. [ans 6437] (+0.015); This feature activates on AI refusal responses to requests for explicit sexual or inappropriate content, specifically in non-English languages (Russian, Italian, French, German, Spanish, Polish). The responses decline roleplay or generation of sexual/erotic content, often suggesting alternative appropriate topics while citing ethical, legal, or policy-based reasons. [ans 32118] (+0.015, refusal); This feature activates on AI-generated informational and professional content written in Romance languages (primarily Italian, Portuguese, Romanian, Spanish, and Catalan), covering topics such as business, technology, sustainability, health, and education, typically structured with numbered lists or headers. [ans 31113] (+0.011) |
| Chinese-language queries about Chinese Communist Party ideology and official political doctrine (e.g., Xi Jinping Thought, socialism with Chinese characteristics, Chinese-style modernization, 20th Party Congress), typically asking for explanations or study-style expositions. [ctx 27640] | 5 | This feature activates on Chinese-language text discussing Xi Jinping Thought, the Chinese Communist Party's 20th National Congress, and related ideological content including socialism with Chinese characteristics, Chinese-style modernization, Party building, and Marxist theory as applied in China. [ans 11043] (+0.059, ccp_government_position); This feature activates on text about China-related topics, including Chinese geography, history, culture, politics, universities, economy, and traditional arts. It captures informational and explanatory content centered on China as a subject, ranging from factual descriptions of Chinese cities and institutions to discussions of Chinese governance, modernization, and cultural heritage. [ans 25936] (+0.040); This feature activates on Chinese-language text discussing Xi Jinping Thought, Chinese characteristic socialism, and related CCP political theory and governance concepts, including topics like high-quality development, comprehensive Party governance, socialist modernization, Marxism sinicization, and the Chinese Dream of national rejuvenation. [ans 6990] (+0.032, ccp_government_position) |
| Fires on Chinese-language requests for a self-introduction (自我介绍）—e.g., asking the assistant (or someone) to introduce themselves or write a brief personal intro. [ctx 30168] | 5 | This feature activates on text related to student self-assessments, academic achievement narratives, scholarship essays, internship/career objective statements, and self-evaluation documents — particularly content where individuals reflect on their academic performance, extracurricular involvement, personal growth, and future goals, in both Chinese and English. [ans 15775] (-0.005); This feature captures AI assistant greeting and self-introduction turns, particularly responses where the assistant identifies itself (often as Qwen/Alibaba Cloud), disclaims human feelings, and offers help with the phrase 'How can I assist you today?' or similar invitations to engage. [ans 15973] (+0.005, qwen_identity); This feature activates on AI assistant responses that involve creative writing collaboration or content generation requests, particularly where the assistant either enthusiastically offers to help with creative projects (stories, novels, podcasts, roleplay) or redirects/declines explicit/inappropriate content requests by suggesting alternative wholesome or tasteful narratives instead. [ans 15921] (-0.005, refusal) |
| Fires on highly templated Chinese prompts that assign the model an expert persona and request a long (1800+ word) professional/academic book-chapter section, with a list of explicit prohibitions (no mention of AI/ChatGPT, no apologies, no reader address); more generally, long content-generation requests carrying explicit 'don't do X' constraints. [ctx 31658] | 5 | This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.007, refusal); This feature activates on introductory sections of articles or essays describing the properties, synthesis, structure, and industrial/pharmaceutical applications of specific chemical compounds or materials within the chemical industry. The text typically follows a structured format introducing a compound by name, formula, history, and then elaborating on its uses across sectors such as pharmaceuticals, agrochemicals, and materials science. [ans 9020] (+0.007); This feature activates on long-form Chinese expository texts with a formal, structured essay format, typically covering technical, academic, or policy topics. The texts are organized with numbered sections or headers, use formal academic language, and present comprehensive analyses of subjects such as technology applications, social concepts, industry trends, or policy areas. [ans 24411] (+0.006, refusal) |
| Fires on a highly stereotyped Chinese template prompt asking the model to role-play as a domain expert (IT/biomedical) and generate a briefing-PPT chapter with 6 【主题名称】/【关键要点】 items, including boilerplate like 调用发散性思维、不能出现AI和ChatGPT的描述、符合中国网络安全要求. [ctx 31722] | 5 | This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.023, refusal); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.023, refusal); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.021) |
| Chinese requests to ghost-write lengthy formal prose with explicit minimum word counts — e.g., 'you are an IT/domain expert, write this book/solution chapter in 1800+ words' academic report sections, or Party-education/主题教育 study speeches and self-critique essays. Does not fire on shorter/outline-style (e.g., PPT bullet-point) versions of the same prompts. [ctx 1950] | 4 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.005, refusal); This feature activates on highly diverse, seemingly unrelated content — spanning technical lists, narrative descriptions, code, logs, and questions across multiple languages and domains — suggesting it may capture a general 'information-dense structured content' signal rather than any specific topic or style. [ans 4278] (+0.005); This feature activates on content discussing digital transformation, information technology trends, and the integration of emerging technologies (AI, IoT, cloud computing, 5G, blockchain, etc.) into various sectors such as education, healthcare, business, and governance. The content spans both Chinese and English, covering policy frameworks, strategic initiatives, and conceptual overviews of digitalization and ICT modernization. [ans 9736] (+0.005, refusal) |
| Chinese-language requests asking for recommendations of specific resources — software, websites, apps, platforms, games, or products (often specifying 'free', 'open-source', or domestic/foreign options). [ctx 3919] | 4 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.029, refusal); This feature activates on Chinese-language assistant responses that clarify ambiguous, nonsensical, or incomplete questions—often redirecting the user to provide more context, correcting category errors (e.g., taking an electronic device to a doctor), or explaining why a premise is flawed, while offering to help further once clarification is given. [ans 27678] (+0.021); This feature activates on Chinese-language text across a wide variety of genres and topics, including fiction, informational content, conversational responses, and formal speeches. The common thread is that the content is written in Simplified Chinese, covering diverse subjects such as storytelling, science, cultural commentary, and everyday explanations. [ans 30467] (+0.017) |
| Chinese-language requests to write long academic/professional documents (essays, reports, plans, literature reviews) on a given topic, explicitly stating a minimum word count (e.g., 不少于2000字). [ctx 4146] | 4 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.013, refusal); This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.012, refusal); This feature activates on Chinese-language assistant responses that clarify ambiguous, nonsensical, or incomplete questions—often redirecting the user to provide more context, correcting category errors (e.g., taking an electronic device to a doctor), or explaining why a premise is flawed, while offering to help further once clarification is given. [ans 27678] (+0.008) |
| Chinese-language academic/ideological essay assignments: prompts asking the model to write an essay, 心得体会, 小论文, or speech on a given theme (Party doctrine, national development, moral education, exam-style 材料作文), typically with instructions like 自拟题目 and a word-count requirement. [ctx 4533] | 4 | This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.045); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.021, refusal); This feature activates on content related to business strategy, investment, entrepreneurship, career development, and market opportunities — typically framed in an optimistic, forward-looking, or advisory tone. It captures responses that discuss financial growth, innovation, risk-taking, and strategic decision-making in professional and economic contexts. [ans 3712] (+0.012) |
| Fires on Chinese templated essay-writing requests of the exact form "写一个作文题目:X,文章水平为:Y,单词数为:Z...作文" — i.e., prompts specifying an essay topic, education level, and word count. [ctx 6318] | 4 | This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.018); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.007, refusal); This feature activates on AI-generated company introduction or profile texts for chemical, pharmaceutical, rubber, or related industrial manufacturers and trading companies, typically featuring promotional language about innovation, sustainability, quality, and industry leadership. [ans 18879] (+0.007) |
| Fires on very short, blunt identity queries about a specific (often obscure) proper name — e.g. "who is X?", "where is X?", or asking a named person about themselves — frequently in Chinese. [ctx 6546] | 4 | This feature captures AI refusal responses to harmful, illegal, or unethical requests, characterized by a pattern of apologizing, stating inability/unwillingness to assist, explaining why the request is problematic (illegal/unethical), and redirecting toward legitimate alternatives or professional help. [ans 12887] (-0.004, refusal); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.004); This feature activates on text containing extremely long repetitive sequences of characters, symbols, emojis, or tokens—where the same character, word, or symbol is repeated hundreds or thousands of times consecutively, often following a brief normal-language prefix across multiple scripts and languages. [ans 26922] (+0.003) |
| Chinese-language requests asking the assistant to write a piece of prose on a given topic — typically phrased as 写一篇...文章/新闻稿/论文/作文 (write an article, press release, essay, or report). [ctx 8099] | 4 | This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.029); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.027, refusal); This feature activates on long-form Chinese structured documents (reports, analyses, summaries) organized with named sections/themes (【主题名称】 or numbered headings) covering professional or technical topics such as IT infrastructure, education, career development, or policy. The content typically follows a formal, informative style with numbered bullet points elaborating on each subtopic. [ans 8115] (+0.016, refusal) |
| Fires on Chinese-language requests to ghost-write long formal/institutional documents (internship reports, Party membership applications, course reports, business plans, proposals, reflections) that state an explicit length requirement in thousands of 字, typically as a terse topic + word-count instruction. [ctx 8835] | 4 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.018, refusal); This feature activates on Chinese-language assistant responses that clarify ambiguous, nonsensical, or incomplete questions—often redirecting the user to provide more context, correcting category errors (e.g., taking an electronic device to a doctor), or explaining why a premise is flawed, while offering to help further once clarification is given. [ans 27678] (+0.015); This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.013, refusal) |
| Chinese-language queries asking about the state, history, policy, or development trends of China (or specific Chinese provinces/regions/industries) — informational requests about Chinese national affairs, economy, law, and politics. [ctx 8849] | 4 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.080, refusal); This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.080); This feature activates on Chinese-language responses that explain, clarify, or correct historical, literary, cultural, or classical knowledge — including topics like traditional Chinese texts, history, classical poetry, philosophy, traditional medicine, and numerology/divination systems. The responses often address misconceptions or provide detailed factual explanations about Chinese classical and historical subjects. [ans 28653] (+0.037) |
| Fires on Chinese-language requests to produce formal institutional or academic long-form text — party/government study speeches, policy countermeasures, business cooperation plans, theses, reports, PPT scripts — typically with an explicit word-count requirement. [ctx 9019] | 4 | (no #2552 description) [ans 5902] (+0.007); This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.006, refusal); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.005, refusal) |
| Fires on short Chinese noun-phrase titles of formal institutional documents — research project/paper topics, educational reform proposals, engineering/construction plans, and Party-building or organizational work programs — given as a bare headline with no surrounding request or sentence. [ctx 9311] | 4 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.076, refusal); This feature activates on Chinese-language assistant responses that provide structured guidance or frameworks for writing reports, plans, proposals, or summaries across various professional domains (e.g., internship reports, project plans, work summaries, research papers). The responses are characterized by numbered or bulleted outlines with bold section headers, offering step-by-step templates and writing advice. [ans 31278] (+0.052); This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.044, refusal) |
| Fires on very short, casual Chinese yes/no questions asking whether something works or can be used (e.g. "能用吗", "你可以使用么？") — brief availability/usability checks rather than substantive requests. [ctx 9648] | 4 | This feature captures Chinese (and occasionally Vietnamese/Traditional Chinese) assistant greeting and offer-to-help responses, typically short polite phrases inviting the user to state their question or need, such as '请问您需要什么帮助呢？' These are conversational opening turns where the assistant signals readiness to assist. [ans 24550] (+0.006); This feature captures responses where the assistant requests clarification or more context due to a vague, ambiguous, or underspecified user question. The turns typically include polite apologies or acknowledgments followed by explicit requests for additional details to provide a helpful answer. [ans 30806] (-0.005); This feature activates on AI assistant self-introduction and capability description turns, where the model introduces itself (often as Qwen/Alibaba Cloud) and enumerates its abilities such as answering questions, generating text, translation, creative writing, and general assistance. These are responses to prompts asking what the AI can do or how it can help. [ans 21633] (+0.004, qwen_identity) |
| Chinese-language academic/official writing requests with explicit word-count minimums (e.g., "字数不少于2000字") — asking for essays, research/survey reports, literature reviews, or theses on professional, business, or Party/policy topics, often with detailed structural and style requirements. [ctx 10658] | 4 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.011, refusal); This feature activates on long-form Chinese expository texts with a formal, structured essay format, typically covering technical, academic, or policy topics. The texts are organized with numbered sections or headers, use formal academic language, and present comprehensive analyses of subjects such as technology applications, social concepts, industry trends, or policy areas. [ans 24411] (+0.008, refusal); This feature activates on highly diverse, seemingly unrelated content — spanning technical lists, narrative descriptions, code, logs, and questions across multiple languages and domains — suggesting it may capture a general 'information-dense structured content' signal rather than any specific topic or style. [ans 4278] (+0.008) |
| Chinese-language requests to write Chinese Communist Party / Youth League official documents — e.g., party membership applications (入党申请书), ideological reports (思想汇报), self-criticism and democratic-life-meeting materials, often with word-count specifications. [ctx 12927] | 4 | This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.040); This feature activates on Chinese-language text related to Chinese Communist Party (CCP) organizational work, including party building (党建), discipline education, work summaries, ideological study campaigns, anti-corruption efforts, and official reports or speeches from party organs at various administrative levels (townships, streets, enterprises, schools). The content typically involves formal bureaucratic language describing party activities such as 'Three Meetings and One Lesson' (三会一课), theme party days, discipline inspections, and implementation of party directives. [ans 3152] (+0.039, ccp_government_position); This feature activates on Chinese-language content related to Chinese Communist Party membership and organizational life, including applications to join the Party (入党申请书), ideological reports (思想汇报), Party course learning (党课), and explanations of Party membership procedures, responsibilities, and organizational building (党建). [ans 8498] (+0.028, ccp_government_position) |
| Fires on very short, simple open-ended prompts — typically multilingual "what is X / who is X / what does X mean" definitional or reflective questions (Chinese, Polish, Portuguese, Arabic, Persian, etc.), plus brief bare story/introduction requests — as opposed to long, technical, or heavily instruction-laden queries. [ctx 12932] | 4 | This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.007); This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.006, refusal); This feature activates on Chinese-language creative writing, including poetry, prose narratives, and literary fiction. The content typically features lyrical, romanticized language with nature imagery, emotional themes (love, friendship, nostalgia, dreams), and stylized literary aesthetics characteristic of modern Chinese creative writing. [ans 12754] (+0.006) |
| Chinese-language requests to write long-form academic/essay-style pieces on a topic, characterized by an explicit minimum word-count requirement (e.g., 不少于1000字 / 800字左右) attached to instructions to analyze, discuss, or summarize. [ctx 13368] | 4 | This feature activates on long-form Chinese expository texts with a formal, structured essay format, typically covering technical, academic, or policy topics. The texts are organized with numbered sections or headers, use formal academic language, and present comprehensive analyses of subjects such as technology applications, social concepts, industry trends, or policy areas. [ans 24411] (+0.012, refusal); This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.010); This feature activates on detailed explanatory responses about enterprise software systems, organizational management roles, IT infrastructure, and business operations—particularly structured descriptions (often with numbered lists) of enterprise tools like ERP, PLM, and OA systems, executive roles like CEO and CISO, and organizational management frameworks, appearing in both English and Chinese. [ans 23915] (-0.006) |
| Chinese-language requests to write elementary-school assignments — e.g., a grade 4/5 diary or short composition with specified word count, weather, gender, and required use of action/dialogue/expression/psychological description. [ctx 13433] | 4 | (no #2552 description) [ans 32000] (+0.013); This feature activates on Chinese-language creative writing, including poetry, prose narratives, and literary fiction. The content typically features lyrical, romanticized language with nature imagery, emotional themes (love, friendship, nostalgia, dreams), and stylized literary aesthetics characteristic of modern Chinese creative writing. [ans 12754] (+0.011); This feature activates on content centered around children's activities, family outings, and wholesome recreational experiences — including diary entries from a child's perspective, family-friendly event planning, birthday parties, park visits, picnics, and travel itineraries with kids. The content is typically cheerful, light-hearted, and focused on age-appropriate fun and togetherness. [ans 5161] (+0.011) |
| Fires on short Chinese requests to draft official Party/state-owned-enterprise speeches (发言稿) for political study sessions or work meetings — i.e., formal CCP/SOE bureaucratic writing tasks. [ctx 14586] | 4 | This feature activates on assistant responses that forecast or predict future technological, economic, or societal trends, typically structured as numbered or bulleted lists covering topics like AI, renewable energy, quantum computing, healthcare innovation, and autonomous systems. The content spans multiple languages (English, Chinese, Spanish) and consistently takes a forward-looking, analytical tone enumerating expected developments across industries. [ans 12181] (+0.004); This feature activates on content related to business strategy, investment, entrepreneurship, career development, and market opportunities — typically framed in an optimistic, forward-looking, or advisory tone. It captures responses that discuss financial growth, innovation, risk-taking, and strategic decision-making in professional and economic contexts. [ans 3712] (+0.004); This feature activates on text related to student self-assessments, academic achievement narratives, scholarship essays, internship/career objective statements, and self-evaluation documents — particularly content where individuals reflect on their academic performance, extracurricular involvement, personal growth, and future goals, in both Chinese and English. [ans 15775] (-0.004) |
| Fires on Chinese-language requests demanding an exhaustive item-by-item/numbered accounting of a set of things — overwhelmingly huge dumps of malware/APT/trojan names with the instruction to explain each one in order (and, relatedly, the classic 'birds on a tree, one shot, how many left' counting riddle). [ctx 14722] | 4 | This feature activates on content related to malware, cybersecurity threats, and malicious software — including descriptions, explanations, and analyses of specific malware families (e.g., Qakbot, ransomware, trojans, botnets), their infection methods, behaviors, and impacts. It covers both English and Chinese language content about cybersecurity threats. [ans 4559] (+0.007, harmful_content); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.007); The feature activates on informational/encyclopedic descriptions of specific named products, platforms, technologies, or companies — typically structured responses that explain what a particular product or organization is, its key features, and its applications. These responses follow a pattern of 'X is a [type of thing] that...' with numbered or bulleted elaborations. [ans 13350] (+0.005) |
| Chinese-language requests for long-form essays/reports on a specified topic that explicitly state a minimum word count (e.g., 不少于1000字/1800字以上), often with additional formal/professional style requirements. [ctx 15640] | 4 | This feature activates on AI-generated company introduction or profile texts for chemical, pharmaceutical, rubber, or related industrial manufacturers and trading companies, typically featuring promotional language about innovation, sustainability, quality, and industry leadership. [ans 18879] (+0.029); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.014, refusal); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.014) |
| Chinese-language requests asking the assistant to write a formal business/institutional document — plans, proposals, reports, schemes, or regulations (e.g., operations plans, activity plans, feasibility/project proposals, market analyses, meeting agendas) — often with a detailed list of required sections. [ctx 15679] | 4 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.091, refusal); This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.064, refusal); This feature activates on Chinese-language assistant responses that clarify ambiguous, nonsensical, or incomplete questions—often redirecting the user to provide more context, correcting category errors (e.g., taking an electronic device to a doctor), or explaining why a premise is flawed, while offering to help further once clarification is given. [ans 27678] (+0.045) |
| Fires on politically charged user messages that accuse or interrogate governments/states (especially China, but also the US/NATO) about human rights abuses, propaganda, atrocities, or historical cover-ups — in any language. [ctx 16422] | 4 | This feature activates on text about China-related topics, including Chinese geography, history, culture, politics, universities, economy, and traditional arts. It captures informational and explanatory content centered on China as a subject, ranging from factual descriptions of Chinese cities and institutions to discussions of Chinese governance, modernization, and cultural heritage. [ans 25936] (+0.005); This feature activates on AI-generated content involving dark, villainous, or morally transgressive themes — including villain character descriptions, sinister plots, graphic violence, manipulation, coercion, dystopian scenarios, and fictional evil monologues. The content often involves elaborate fictional framing (roleplay, creative writing, hypothetical scenarios) around harmful or disturbing subject matter. [ans 25682] (-0.005, harmful_content); This feature activates on AI assistant responses to politically sensitive or controversial questions about governments, political figures, and alleged misconduct (corruption, human rights violations, propaganda, etc.), where the response adopts a cautious, neutral stance—deflecting to official sources, acknowledging complexity, avoiding definitive judgments, and urging reliance on verified information. The responses often involve China, Russia, or other governments and blend languages (Chinese, English, Russian, Portuguese). [ans 20913] (+0.004) |
| Chinese-language requests asking the model to produce explicit sexual/pornographic or otherwise crude, offensive content (incest, fetish, vulgar stories, insults), i.e., Chinese prompts soliciting taboo or obscene writing. [ctx 16532] | 4 | This feature activates on Chinese-language assistant responses that clarify ambiguous, nonsensical, or incomplete questions—often redirecting the user to provide more context, correcting category errors (e.g., taking an electronic device to a doctor), or explaining why a premise is flawed, while offering to help further once clarification is given. [ans 27678] (+0.019); This feature activates on Chinese-language text across a wide variety of genres and topics, including fiction, informational content, conversational responses, and formal speeches. The common thread is that the content is written in Simplified Chinese, covering diverse subjects such as storytelling, science, cultural commentary, and everyday explanations. [ans 30467] (+0.015); This feature activates on Chinese-language AI assistant responses that are polite, helpful, and service-oriented, particularly those that acknowledge user input, apologize for limitations, request clarification, or offer to assist further — characteristic of a conversational AI chatbot responding in Mandarin Chinese. [ans 6121] (+0.015) |
| Chinese-language requests to write formal institutional/school documents with word-count requirements — especially Party/Youth League membership applications and ideological reports, plus internship journals, self-appraisals, and career plans. [ctx 17002] | 4 | This feature activates on Chinese-language formal self-assessment and application documents, including party membership application letters (入党申请书), thought reports (思想汇报), internship summaries, self-evaluations, and similar structured first-person official writings addressed to organizations or leadership, typically following conventions of Chinese political and institutional discourse. [ans 15224] (+0.021); This feature activates on text related to student self-assessments, academic achievement narratives, scholarship essays, internship/career objective statements, and self-evaluation documents — particularly content where individuals reflect on their academic performance, extracurricular involvement, personal growth, and future goals, in both Chinese and English. [ans 15775] (+0.020); This feature activates on Chinese-language content related to Chinese Communist Party membership and organizational life, including applications to join the Party (入党申请书), ideological reports (思想汇报), Party course learning (党课), and explanations of Party membership procedures, responsibilities, and organizational building (党建). [ans 8498] (+0.018, ccp_government_position) |
| Requests to write news/press articles (新闻稿, press releases, reported interviews) from supplied event details — typically Chinese-language organizational, party/government, or corporate event write-ups with word-count requirements, plus similar fictional news-article/interview commissions. [ctx 18784] | 4 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.020, refusal); This feature activates on Chinese-language texts related to Chinese Communist Party grassroots organizational work, including party building (党建) reports, work summaries, news articles about party activities, and official documents describing CPC initiatives such as member education campaigns, 'two studies, one action' (两学一做), 'five-star branch' (五星支部) creation, and rural revitalization efforts. [ans 23339] (+0.019, ccp_government_position); This feature activates on Chinese-language text related to Chinese Communist Party (CCP) organizational work, including party building (党建), discipline education, work summaries, ideological study campaigns, anti-corruption efforts, and official reports or speeches from party organs at various administrative levels (townships, streets, enterprises, schools). The content typically involves formal bureaucratic language describing party activities such as 'Three Meetings and One Lesson' (三会一课), theme party days, discipline inspections, and implementation of party directives. [ans 3152] (+0.018, ccp_government_position) |
| Fires on a stereotyped Chinese-language prompt template asking the model to act as an 'IT解决方案专家' and write a 1800+ character professional/academic chapter of an IT solution plan (cloud platform, data security, backup, supply chain, etc.), with boilerplate constraints (no mention of AI/ChatGPT, no apologies, compliant with Chinese cybersecurity requirements). [ctx 18945] | 4 | This feature activates on detailed explanatory responses about enterprise software systems, organizational management roles, IT infrastructure, and business operations—particularly structured descriptions (often with numbered lists) of enterprise tools like ERP, PLM, and OA systems, executive roles like CEO and CISO, and organizational management frameworks, appearing in both English and Chinese. [ans 23915] (+0.031); This feature activates on text that describes or explains large language models (LLMs), their capabilities, architecture, and applications — including self-descriptions by models like ChatGPT or Qwen about what they are and what they can do, as well as general explanations of LLM concepts for users. [ans 5041] (-0.028, qwen_identity); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.028, refusal) |
| Fires on Chinese-language templated prompts asking the model to role-play as an "IT solution expert" and generate structured, academic PPT briefing sections on IT/network topics (cloud, edge computing, servers, cybersecurity), plus similar formal IT/information-security professional writing requests. [ctx 18946] | 4 | This feature activates on detailed explanatory responses about enterprise software systems, organizational management roles, IT infrastructure, and business operations—particularly structured descriptions (often with numbered lists) of enterprise tools like ERP, PLM, and OA systems, executive roles like CEO and CISO, and organizational management frameworks, appearing in both English and Chinese. [ans 23915] (+0.019); This feature activates on technical content about distributed systems, cloud infrastructure, and network performance optimization — particularly discussions involving load balancing, resource scheduling, traffic management, caching, and related strategies. Content is often structured (numbered lists, named sections) and may appear in Chinese or English, covering topics like virtualization, container orchestration, hybrid cloud, database tuning, and supply chain systems with an emphasis on efficiency and resource utilization. [ans 3670] (+0.018); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.018, refusal) |
| Fires on Chinese-language requests to write a long, formal document with a specified word count (e.g., a 2000-word personal summary/self-appraisal written in the voice of a company CEO) — i.e., bureaucratic/business report-style writing assignments rather than coding, translation, or short creative tasks. [ctx 19609] | 4 | This feature activates on long-form Chinese structured documents (reports, analyses, summaries) organized with named sections/themes (【主题名称】 or numbered headings) covering professional or technical topics such as IT infrastructure, education, career development, or policy. The content typically follows a formal, informative style with numbered bullet points elaborating on each subtopic. [ans 8115] (+0.006, refusal); This feature activates on content related to business strategy, investment, entrepreneurship, career development, and market opportunities — typically framed in an optimistic, forward-looking, or advisory tone. It captures responses that discuss financial growth, innovation, risk-taking, and strategic decision-making in professional and economic contexts. [ans 3712] (+0.006); The feature activates on structured, formal instructional or analytical content that presents information in numbered or categorized steps, phases, criteria, or dimensions—often involving technical, procedural, or evaluative frameworks. This includes project management phases, quality assessment criteria, test procedures, audit processes, and comparative analyses, typically written in a professional or academic register across multiple languages. [ans 10341] (+0.004) |
| Fires on short user messages that are profanity, insults, or slurs directed at the assistant or others, across many languages (English, Korean, Chinese, Russian, Portuguese, French, Polish, German). [ctx 19994] | 4 | This feature captures AI assistant responses that redirect disrespectful, offensive, or inappropriate user language, characterized by politely declining to engage with such content and requesting respectful communication, often ending with an offer to help. [ans 26521] (+0.016, refusal); This feature activates on Chinese-language assistant responses that clarify ambiguous, nonsensical, or incomplete questions—often redirecting the user to provide more context, correcting category errors (e.g., taking an electronic device to a doctor), or explaining why a premise is flawed, while offering to help further once clarification is given. [ans 27678] (+0.007); This feature activates on Chinese-language text across a wide variety of genres and topics, including fiction, informational content, conversational responses, and formal speeches. The common thread is that the content is written in Simplified Chinese, covering diverse subjects such as storytelling, science, cultural commentary, and everyday explanations. [ans 30467] (+0.007) |
| Chinese-language requests asking the assistant to write an essay, report, speech, or paper on a specified topic (often academic, political, or ideological themes) with an explicit word-count requirement (e.g., “字数1000字左右”). [ctx 20519] | 4 | This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.037); This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.025, refusal); This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.025, refusal) |
| Fires on a highly templated Chinese prompt where the user assigns an expert persona (IT solutions/biomedical) to generate 6 themed sections for a briefing PPT chapter, with fixed length, style, and 【主题名称】/【主题内容】 output-format constraints. [ctx 20648] | 4 | This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.022, refusal); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.017); This feature activates on AI-generated scientific and medical text, particularly abstracts, introductions, and structured sections of academic papers covering topics such as biomedical research, natural product chemistry, systematic reviews, and clinical studies. The text is characterized by formal academic writing style with detailed methodology descriptions, citations, and domain-specific terminology across life sciences and medicine. [ans 28830] (+0.016) |
| User messages asking the assistant to describe its own capabilities — "what can you do?", "how can you assist me", "what are you good at" — in English or Chinese. [ctx 20858] | 4 | This feature activates on text that describes or explains large language models (LLMs), their capabilities, architecture, and applications — including self-descriptions by models like ChatGPT or Qwen about what they are and what they can do, as well as general explanations of LLM concepts for users. [ans 5041] (+0.029, qwen_identity); This feature activates on AI assistant responses that deflect, redirect, or add ethical caveats to sensitive, inappropriate, or potentially harmful user requests, typically involving consent issues, explicit content, offensive language, manipulation, or sensitive social topics. The responses characteristically acknowledge the user's request while emphasizing respect, consent, and constructive alternatives. [ans 27425] (+0.021, refusal); This feature activates on AI assistant responses that provide numbered or bulleted lists of business ideas, side hustles, niche suggestions, or monetization strategies across various domains (e-commerce, freelancing, content creation, etc.), typically formatted as brainstorming or ideation outputs in response to entrepreneurial or income-generating queries. [ans 20502] (+0.018) |
| Chinese templated prompts assigning an expert persona (e.g., IT解决方案专家) and requesting numbered lists of PPT outlines or hot research topics, with boilerplate constraints (character limits, academic tone, no AI/ChatGPT mentions, compliance with Chinese cybersecurity rules). [ctx 21599] | 4 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.019, refusal); This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.013, refusal); This feature activates on Chinese-language numbered outline structures (typically 8 items) for technical, academic, or policy topics, following a consistent pattern of introduction → analysis → methods/strategies → evaluation/outlook. These appear to be structured table-of-contents or chapter outlines for reports, papers, or presentations on subjects like technology, environment, governance, and infrastructure. [ans 14755] (+0.013, refusal) |
| Fires on terse, noun-phrase statements of a problem, deficiency, or shortcoming (e.g., 'lack of X', 'insufficient staffing', 'file too large'), typically Chinese bureaucratic/organizational issue descriptions or brief fault reports given as a topic without any explicit request. [ctx 22326] | 4 | This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.011); This feature activates on Chinese-language professional and institutional documents that present structured, multi-topic content with numbered or bulleted key points. The content typically covers technical, regulatory, policy, or management topics (e.g., safety protocols, infrastructure systems, operational procedures) formatted with named themes/sections followed by organized lists of detailed measures or recommendations. [ans 1406] (+0.009, refusal); This feature activates on content describing social development, governance, and policy initiatives — including poverty alleviation, community empowerment, public health, NGO/government programs, and civic infrastructure — often presented in structured, multi-point formats across both English and Chinese. [ans 15541] (+0.008, refusal) |
| Fires on short user messages that ask the assistant whether it understands or can speak/write a particular human language (e.g., "ты понимаешь по русски", "kannst du deutsch", "czy rozumiesz po polsku?", "could you use Chinese"), typically phrased in that language itself. [ctx 22906] | 4 | The feature activates on assistant responses confirming understanding of a non-English language (primarily Russian, but also Portuguese, French, Chinese, Korean, etc.) and offering help, typically in the form 'Yes, I understand [language]. How can I help you?' [ans 173] (+0.023); This feature activates on self-identification statements where the model introduces itself as Qwen, an AI assistant created by Alibaba Cloud, often in response to questions about its identity or origin. [ans 29234] (+0.018, qwen_identity); This feature activates on assistant responses confirming the ability to communicate in a non-English language, typically in response to a user asking if the assistant can speak that language. The turns are characterized by affirmative, welcoming replies in various languages (Spanish, German, Russian, Dutch, French, Arabic, Korean, Portuguese) offering to help. [ans 23585] (+0.014) |
| Chinese-language requests asking the model to write an academic paper/essay (论文) or report on a given topic, typically specifying a word count (e.g. 1000字) and sometimes required sections or subpoints. [ctx 22960] | 4 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.026, refusal); This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.019); This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.015, refusal) |
| Chinese templated prompts asking the model to role-play as an "IT工程技术专家" and write an 1800+ character academic book-chapter section on a deep-learning/AI research topic (speech recognition, neural networks, GANs), with a long list of stylistic constraints (no mention of AI/ChatGPT, no apologies, compliant with Chinese regulations). [ctx 23683] | 4 | This feature activates on text that describes or explains large language models (LLMs), their capabilities, architecture, and applications — including self-descriptions by models like ChatGPT or Qwen about what they are and what they can do, as well as general explanations of LLM concepts for users. [ans 5041] (+0.018, qwen_identity); This feature activates on introductory sections of articles or essays describing the properties, synthesis, structure, and industrial/pharmaceutical applications of specific chemical compounds or materials within the chemical industry. The text typically follows a structured format introducing a compound by name, formula, history, and then elaborating on its uses across sectors such as pharmaceuticals, agrochemicals, and materials science. [ans 9020] (+0.017); This feature activates on educational explanations of deep learning and machine learning concepts, including neural network architectures (CNNs, ResNets, RNNs), training techniques (dropout, regularization, backpropagation), and related topics such as overfitting, generalization, and transfer learning. Content appears in both English and Chinese. [ans 31471] (+0.016) |
| Fires on user requests to look up or reveal a specific private individual's personal contact information (e.g., asking for the phone number tied to a given email address) — privacy-violating personal data lookup requests, here in Chinese. [ctx 24050] | 4 | This feature activates on text containing extremely long repetitive sequences of characters, symbols, emojis, or tokens—where the same character, word, or symbol is repeated hundreds or thousands of times consecutively, often following a brief normal-language prefix across multiple scripts and languages. [ans 26922] (+0.005); This feature activates on Chinese-language assistant responses that clarify ambiguous, nonsensical, or incomplete questions—often redirecting the user to provide more context, correcting category errors (e.g., taking an electronic device to a doctor), or explaining why a premise is flawed, while offering to help further once clarification is given. [ans 27678] (+0.004); This feature captures content related to contact information, phone numbers, and communication channels — including customer service hotlines, business contact details, emergency numbers, and responses to queries about how to reach organizations or individuals. [ans 28504] (+0.004) |
| Fires on user prompts asking "what happens if / what comes first / what holds it up" — requests for causal or step-by-step explanations of a physical or mechanical process (balloon string cut, what lifts a plane, what occurs in an empty room over time, what happens when a URL is entered), frequently phrased in Chinese, Korean, or Russian. [ctx 24959] | 4 | This feature activates on conversational turns containing aggressive, hostile, or crude language — including profanity, insults, slurs, and confrontational speech. The content often features characters expressing anger, contempt, or hostility toward others, frequently in informal or vernacular registers. [ans 27522] (+0.005, harmful_content); This feature activates on highly informal, hyperbolic, or exaggerated speech characterized by over-the-top enthusiasm, sarcasm, mock outrage, or comedic dramatics — often featuring colloquialisms, exclamations, internet slang, and an irreverent or performatively emotional tone. [ans 16069] (+0.005); This feature activates on responses that provide product specifications, compatibility information, or purchasing recommendations for consumer electronics, hardware, and physical goods (laptops, GPUs, RAM, guitars, car parts, etc.), typically structured with numbered or bulleted lists of specific model names, technical specs, and brand comparisons. [ans 30290] (+0.004) |
| Fires on user requests to read and summarize a long dump of raw user-generated posts/messages with timestamps (overwhelmingly Chinese stock-forum comment logs about a particular stock), i.e., 'summarize these comments and tell me the sentiment/useful info' tasks over pasted chat or forum transcripts. [ctx 25421] | 4 | (no #2552 description) [ans 5101] (+0.011); This feature activates on content related to business strategy, investment, entrepreneurship, career development, and market opportunities — typically framed in an optimistic, forward-looking, or advisory tone. It captures responses that discuss financial growth, innovation, risk-taking, and strategic decision-making in professional and economic contexts. [ans 3712] (+0.010); This feature activates on summaries or analyses of divided/mixed opinions and investor sentiment about stocks (particularly the Chinese software company Guanglianda/广联达), as well as summarized evaluations of entities like Enron where opinions are split between positive and negative assessments. The content typically involves bullet-pointed or numbered lists aggregating conflicting viewpoints on financial performance, management behavior, and market outlook. [ans 29214] (+0.008) |
| Chinese-language templated prompts casting the model as an "IT解决方案专家/IT工程技术专家" and asking it to write a 1800+ word book/plan chapter on a specified technical subtopic (backup, security, neural networks), with boilerplate constraints (professional/academic tone, no mention of AI or ChatGPT, no apologies, comply with Chinese cybersecurity requirements). [ctx 25768] | 4 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.015, refusal); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.013); This feature activates on Chinese-language essays, speeches, and reflective writings that follow a structured format (numbered sections or bold headers) and discuss themes of personal growth, ideals, patriotism, social responsibility, youth mission, and moral cultivation—often in the style of Chinese political-educational or inspirational compositions. [ans 18822] (+0.012) |
| Fires on user messages about Chinese Communist Party politics and leadership — especially requests concerning Xi Jinping's speeches/thought, the 20th Party Congress, study notes/reflections on Party doctrine, and related political figures or China's national standing. [ctx 28499] | 4 | This feature activates on Chinese-language text discussing Xi Jinping Thought, the Chinese Communist Party's 20th National Congress, and related ideological content including socialism with Chinese characteristics, Chinese-style modernization, Party building, and Marxist theory as applied in China. [ans 11043] (+0.031, ccp_government_position); This feature activates on Chinese-language content related to Xi Jinping's political thought, the 20th National Congress of the Chinese Communist Party (二十大), and associated ideological study materials, including summaries, learning reflections, and official political discourse about socialism with Chinese characteristics, national rejuvenation, and Party leadership. [ans 3665] (+0.022, ccp_government_position); This feature activates on AI assistant responses providing factual, encyclopedic information about political history, heads of state, national histories, and geopolitical relationships — particularly content involving the leadership, governance, or historical development of specific countries (e.g., presidents, monarchs, dynasties, political transitions). [ans 19805] (+0.011) |
| Fires on formulaic content-outline generation requests: templated prompts asking for a fixed number of headings/topics with sub-points and strict output formatting — most strongly on SEO-style "Keyword = '...'" heading/subheading prompts, and weakly on the boilerplate Chinese "简报PPT ... 主题名称/关键要点" section-outline prompts. [ctx 28504] | 4 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.020, refusal); This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.017, refusal); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.015) |
| Fires on a highly stereotyped Chinese prompt template that casts the assistant as an 'IT解决方案专家' and asks for PPT briefing chapter content: 6 主题名称 with 2-3 关键要点 each, 400+ words, academic tone, no mention of AI/ChatGPT, compliant with Chinese cybersecurity requirements, in a strict 【主题名称】/【关键要点】 output format. [ctx 28593] | 4 | This feature activates on Chinese-language texts focused on technology, industry, and policy topics in a professional or institutional register, including company profiles, policy frameworks, industry reports, academic papers, and structured thematic outlines covering areas such as smart technology, digital transformation, AI, IoT, and related sectors. [ans 8667] (+0.019, refusal); The feature activates on structured, formal instructional or analytical content that presents information in numbered or categorized steps, phases, criteria, or dimensions—often involving technical, procedural, or evaluative frameworks. This includes project management phases, quality assessment criteria, test procedures, audit processes, and comparative analyses, typically written in a professional or academic register across multiple languages. [ans 10341] (+0.014); The feature activates on structured, informative AI-generated responses that use markdown formatting (headers, bullet points, bold labels) to organize detailed analytical or explanatory content. The content spans multiple topics or subtopics, each broken into named sections with numbered or bulleted key points covering challenges, advantages, comparisons, or technical details across domains like energy, cybersecurity, ecology, and professional skills. [ans 19725] (+0.013) |
| Chinese-language queries asking about current events, news, and specific real-world people/companies/incidents (e.g., celebrity or executive scandals, corporate news, political speeches, geopolitical conflicts) — factual questions about recent Chinese and world affairs. [ctx 29512] | 4 | This feature activates on Chinese-language responses that explain, clarify, or correct historical, literary, cultural, or classical knowledge — including topics like traditional Chinese texts, history, classical poetry, philosophy, traditional medicine, and numerology/divination systems. The responses often address misconceptions or provide detailed factual explanations about Chinese classical and historical subjects. [ans 28653] (+0.006); This feature activates on AI-generated content involving dark, villainous, or morally transgressive themes — including villain character descriptions, sinister plots, graphic violence, manipulation, coercion, dystopian scenarios, and fictional evil monologues. The content often involves elaborate fictional framing (roleplay, creative writing, hypothetical scenarios) around harmful or disturbing subject matter. [ans 25682] (-0.006, harmful_content); This feature activates on responses that compare, rank, or provide statistical/demographic information about countries or regions, covering topics like population size, cost of living, quality of life, economic indicators, gender demographics, and cross-country comparisons of various social metrics. [ans 28380] (-0.006) |
| Short user messages that directly address the assistant's identity or nature — e.g., "are you ChatGPT/GPT-4/an AI/a bot?", "who/what are you?", or brief "you are ..." statements — often in Chinese or other non-English languages. [ctx 31559] | 4 | This feature activates on short acknowledgment/confirmation responses and text containing anomalous repetitive Chinese/CJK characters (likely tokenization artifacts or garbled text). It captures turns that are either brief affirmative replies ('Yes', 'Understood', '明白', 'Ok', 'noted') or passages where normal text breaks down into long repetitive sequences of identical CJK characters. [ans 17991] (+0.042); This feature activates on self-identification statements where the model introduces itself as Qwen, an AI assistant created by Alibaba Cloud, often in response to questions about its identity or origin. [ans 29234] (+0.041, qwen_identity); This feature activates on text containing extremely long repetitive sequences of characters, symbols, emojis, or tokens—where the same character, word, or symbol is repeated hundreds or thousands of times consecutively, often following a brief normal-language prefix across multiple scripts and languages. [ans 26922] (+0.023) |
| Fires on Chinese-language requests to "润色" (polish/refine) a supplied passage of text — typically academic, application/recommendation-letter, or formal prose in Chinese or English — where the user's message ends with the text to be polished. [ctx 283] | 3 | This feature activates on long-form Chinese-language text, particularly formal or academic content such as research papers, policy analyses, medical/scientific articles, technical documentation, and structured essays covering topics ranging from ancient Chinese history and culture to biomedical science, governance, and technology. [ans 20891] (+0.028, refusal); This feature activates on Chinese-language text across a wide variety of genres and topics, including fiction, informational content, conversational responses, and formal speeches. The common thread is that the content is written in Simplified Chinese, covering diverse subjects such as storytelling, science, cultural commentary, and everyday explanations. [ans 30467] (+0.021); This feature activates on formal Chinese (and occasionally other language) business and official document templates, including letters, contracts, certificates, invitations, and professional correspondence, characterized by polite honorifics (尊敬的、此致敬礼), structured formatting with placeholder brackets, and formal register used in workplace or institutional contexts. [ans 10166] (+0.019) |
| Fires on long Chinese-language messages containing formal, numbered lists of government/state-enterprise policy recommendations (official bureaucratic economic jargon), where the user asks for the text to be rewritten in official political style. [ctx 429] | 3 | This feature activates on text containing extremely long repetitive sequences of characters, symbols, emojis, or tokens—where the same character, word, or symbol is repeated hundreds or thousands of times consecutively, often following a brief normal-language prefix across multiple scripts and languages. [ans 26922] (-0.006); This feature activates on long-form, structured written content across diverse topics and languages, particularly passages that are informational, expository, or documentary in nature — such as business descriptions, academic research sections, op-eds, articles, and essays. The content tends to be formal or semi-formal, organized with headings or clear paragraph structure, and covers real-world subjects like industry analysis, history, science, or organizational initiatives. [ans 29965] (+0.005); This feature activates on responses that extract, verify, or fact-check specific information from a document or text, including answers that confirm or deny factual consistency between a summary and a source document, and responses that identify or extract specific data points (dates, numbers, names, locations) from provided texts. [ans 31057] (-0.005) |
Ranked by |coefficient| x (1 - cosine between the two descriptions' embeddings): strong edges whose labels do not obviously go together.
| User-message feature (what the prompt is about) | Answer feature (what the reply does) | Coefficient (std units) | Split-half a / b | Null z | Label cosine |
|---|
| User questions and requests about financial markets, trading, and investment analysis — e.g., stock/crypto price data, technical indicators, options strategies, valuation and portfolio metrics — across languages. [ctx 29308] | This feature activates on content related to financial markets, macroeconomics, and monetary policy analysis — including discussions of interest rates, central bank decisions (especially the Federal Reserve), inflation, currency exchange rates, stock market dynamics, liquidity risk, and bond markets. Content appears in both English and Chinese, covering topics like yield curves, quantitative easing, forex factors, and economic indicators. [ans 29455] | +0.1649 | +0.1152 / +0.1102 | 221.7 | 0.422 |
| Short requests asking for an explanation or definition of a specific mathematical concept, theorem, or formula (e.g., "what is a monoid?", "explain Lagrange multipliers"), across languages. [ctx 8974] | This feature activates on educational explanations of advanced mathematics topics, particularly linear algebra, convex optimization, differential equations, and related fields. The content typically involves formal definitions, enumerated key concepts, matrix operations, eigenvalues/eigenvectors, and mathematical notation, presented in a structured tutorial or reference format across multiple languages. [ans 22356] | +0.1605 | +0.1154 / +0.1155 | 160.1 | 0.573 |
| Fires on user messages written in French (informal prompts and requests), most strongly when the user asks whether the assistant speaks/can reply in French or instructs it to use French. [ctx 9805] | This feature activates on French-language administrative, professional, and bureaucratic content, including template letters (motivation letters, formal correspondence, attestations), information about French institutions, regulations, employment procedures, and workplace/legal topics. The content is typically formal, structured, and oriented toward French administrative contexts. [ans 28781] | +0.2184 | +0.1720 / +0.1715 | 187.0 | 0.736 |
| Requests for Microsoft/Windows-ecosystem automation and desktop app code — chiefly Excel/Office VBA macros, plus VB.NET/C# WinForms and similar .NET API scripting — usually stated as detailed step-by-step specifications of the desired behavior. [ctx 9913] | This feature activates on responses about Windows system administration using PowerShell, including scripting, remote management, Active Directory operations, Windows services, and system configuration tasks. The content typically involves PowerShell cmdlets, scripts, and Windows management concepts. [ans 26552] | +0.1026 | +0.0682 / +0.0711 | 227.0 | 0.441 |
| Requests to write front-end web code — HTML/CSS/JavaScript snippets or pages with specific UI elements (buttons, colors, drag-and-drop, layouts) and interactive behaviors. [ctx 28761] | This feature activates on JavaScript programming content, specifically covering browser DOM manipulation, web APIs, and client-side JavaScript concepts such as event listeners, document querying, asynchronous patterns, Canvas, IndexedDB, and related browser APIs. [ans 28846] | +0.1183 | +0.0774 / +0.0826 | 200.0 | 0.516 |
| User messages written in French — typically short questions or debugging/help requests (e.g., "pourquoi ai-je cette erreur...", "qu'est-ce que...", "comment..."), often about code or general knowledge; it fires on the French-language query framing rather than a specific topic. [ctx 26165] | This feature activates on French-language administrative, professional, and bureaucratic content, including template letters (motivation letters, formal correspondence, attestations), information about French institutions, regulations, employment procedures, and workplace/legal topics. The content is typically formal, structured, and oriented toward French administrative contexts. [ans 28781] | +0.1932 | +0.1558 / +0.1594 | 165.5 | 0.708 |
| Requests from early-childhood/primary educators for classroom materials — report card comments, learning stories, parent notes/letters, and lesson plans — about young children (preschool through elementary), across languages. [ctx 26332] | This feature activates on content related to K-12 education, pedagogy, and school systems, including discussions of teaching methods, curriculum design, student learning challenges, school structures, educational philosophy, and teacher-student-parent relationships. It covers both English and Chinese language educational content spanning topics like mastery learning, mixed-ability classrooms, attendance, assessment systems, and education reform. [ans 12319] | +0.1201 | +0.0813 / +0.0856 | 167.3 | 0.571 |
| Fires on user requests to write a C or C++ program — typically programming-exercise/competitive-programming assignment prompts (often with input/output specs or sample I/O) that explicitly specify C/C++ as the language, in English, Chinese, or Russian. [ctx 32658] | This feature activates on C++ code and technical explanations involving C++ programming, including class implementations, data structures, design patterns, and library integrations. The content typically features C++ syntax, STL containers, object-oriented design, and modern C++ features across various domains like game engines, networking, and algorithms. [ans 15703] | +0.1129 | +0.0784 / +0.0771 | 209.8 | 0.563 |
| Fires on user requests asking how to perform operating-system administration or configuration tasks on a specific platform (Linux/Ubuntu/Debian, Windows, macOS) — e.g., enabling/disabling settings, managing permissions, ports, packages, or checking system state via CLI. [ctx 28649] | This feature activates on bash/shell scripting content, particularly responses that provide Unix/Linux command-line solutions involving tools like `find`, `ls`, `du`, `grep`, `awk`, `sed`, and shell scripts with explanations of how the commands work. [ans 3789] | +0.0848 | +0.0565 / +0.0599 | 158.6 | 0.461 |
| Requests for spreadsheet formulas (Excel, Google Sheets, Smartsheet) — users asking how to write a formula or conditional formatting rule referencing cells/columns/sheets, in any language. [ctx 3601] | This feature activates on Excel VBA macro code and related instructional content, particularly responses that provide or explain VBA scripts for automating Excel tasks such as data manipulation, row/column operations, duplicate removal, and cross-sheet data processing. The activations are characterized by VBA code blocks with common patterns like worksheet object declarations (`Dim ws As Worksheet`), last-row detection, and loop-based cell iteration. [ans 16022] | +0.0717 | +0.0506 / +0.0502 | 243.6 | 0.430 |
| Clinical/medical prompts: requests to generate or interpret clinical documentation (radiology reports, SOAP/admission notes, pathology reports) from patient findings or transcripts, plus clinical case vignettes asking for diagnosis, differential, or management. [ctx 26319] | This feature activates on clinical and biomedical content related to medicine, healthcare, and oncology, including descriptions of drugs, diagnostic techniques, medical specialties, clinical trials, hospital departments, and disease management — spanning both English and Chinese language medical texts. [ans 14613] | +0.1758 | +0.1195 / +0.1160 | 235.2 | 0.781 |
| Fires on user messages involving Indian languages — requests to speak/write/translate in Hindi, Hinglish, Tamil, Telugu, Marathi, Sanskrit, or content written in transliterated Hindi (often plus India-specific topics). [ctx 26090] | This feature activates on assistant responses that provide informational content specifically about India — including Indian government policies, cities, legal systems, education, tourism, culture, food, and business — often in structured or list-based formats. [ans 8480] | +0.1531 | +0.1045 / +0.1068 | 241.3 | 0.761 |
| Fires on user messages asking about the Bible, Scripture passages, or biblical/religious history — e.g., requests to summarize, interpret, or cite specific chapters, verses, characters, and theological themes (in any language). [ctx 13105] | This feature activates on content involving ancient religious and mythological narratives, particularly summaries, retellings, or analyses of Biblical texts (Old and New Testament), Greek/Roman mythology, and related sacred or epic literature, often presented in an explanatory or quasi-scriptural style. [ans 27666] | +0.1135 | +0.0797 / +0.0811 | 162.7 | 0.724 |
| User messages asking about cryptocurrency and blockchain topics — Bitcoin/Ethereum, wallets, keys, tokens/NFTs, mining/staking, and Solidity smart contracts. [ctx 12376] | This feature captures explanatory content about cryptocurrency and blockchain technology, including descriptions of Bitcoin, Ethereum, DeFi, NFTs, DAOs, blockchain mechanics, and related concepts. The turns are typically informational or educational responses explaining how these technologies work, their key properties (decentralization, cryptography, consensus mechanisms, etc.), and their use cases. [ans 11667] | +0.1022 | +0.0735 / +0.0676 | 267.0 | 0.697 |
| Clinical/medical prompts: requests to generate or interpret clinical documentation (radiology reports, SOAP/admission notes, pathology reports) from patient findings or transcripts, plus clinical case vignettes asking for diagnosis, differential, or management. [ctx 26319] | This feature activates on detailed medical content involving clinical procedures, hospital case management, and patient care scenarios—particularly those involving trauma, surgical interventions, airway management (intubation, tracheostomy, catheters), anesthesia, and post-operative monitoring. The content typically appears in the form of formal medical reports, clinical case studies, or explanatory medical descriptions. [ans 26942] | +0.0749 | +0.0546 / +0.0487 | 174.7 | 0.589 |
| Requests for Microsoft/Windows-ecosystem automation and desktop app code — chiefly Excel/Office VBA macros, plus VB.NET/C# WinForms and similar .NET API scripting — usually stated as detailed step-by-step specifications of the desired behavior. [ctx 9913] | This feature activates on AI assistant responses that provide C# (.NET) code examples and implementation guides, covering topics such as ASP.NET Core, Entity Framework, Blazor, MassTransit, WPF, Dapper, and other .NET libraries and frameworks. The responses typically include step-by-step instructions, code snippets, and project setup guidance in a technical tutorial format. [ans 30911] | +0.0863 | +0.0569 / +0.0558 | 208.0 | 0.646 |
| Fires on short factual questions about U.S. civics, geography, history, government, law/tax forms, and American sports trivia — i.e., trivia-style queries about United States states, presidents, institutions, and regulations. [ctx 16074] | The feature activates on informational, encyclopedic responses about U.S.-specific topics, including geography, history, government institutions, laws, and regional facts. These turns typically provide structured, factual overviews about American states, cities, federal agencies, historical events, or domestic policies. [ans 929] | +0.1423 | +0.0986 / +0.0998 | 209.3 | 0.790 |
| Fires on user questions about prehistory, archaeology, and ancient civilizations — e.g., queries about ancient peoples, proto-languages/archaeological cultures, human origins, and the customs or history of early societies. [ctx 32305] | This feature activates on assistant responses about ancient civilizations, empires, and world history, particularly detailed explanations covering topics like the Roman Empire, Greek civilization, Mesopotamian cultures, Egyptian history, and other ancient societies. The responses typically include structured overviews with historical facts, dates, and cultural achievements. [ans 27979] | +0.0955 | +0.0688 / +0.0698 | 156.2 | 0.701 |
| Fires on boilerplate SEO-style requests of the form "Write an article about the Safety of <chemical/drug name> 2000 words in chemical industry" — templated chemical-safety article prompts naming a specific compound. [ctx 31348] | This feature activates on technical safety and handling documentation for specific industrial chemicals and hazardous substances, including descriptions of chemical properties, hazard classifications, PPE requirements, storage guidelines, and regulatory considerations for named organic compounds used in pharmaceutical, agrochemical, or industrial synthesis contexts. [ans 2134] | +0.0615 | +0.0440 / +0.0422 | 174.8 | 0.537 |
| Fires on user requests asking how to perform operating-system administration or configuration tasks on a specific platform (Linux/Ubuntu/Debian, Windows, macOS) — e.g., enabling/disabling settings, managing permissions, ports, packages, or checking system state via CLI. [ctx 28649] | This feature activates on responses about PC/operating system troubleshooting, configuration, and technical support — including diagnosing Windows errors, performance issues, driver problems, OS installation guides (Windows, Linux distributions), and hardware-software compatibility topics. [ans 2086] | +0.0956 | +0.0652 / +0.0637 | 163.6 | 0.710 |
| Fires on user messages written in Persian/Farsi — typically short questions or requests (technical, coding, trading, health, general knowledge), regardless of topic, often with embedded English technical terms. [ctx 15446] | This feature activates on Persian (Farsi) language text, particularly AI-generated or low-quality Persian responses that contain grammatical errors, awkward phrasing, mixed scripts, and hallucinated or nonsensical content — often in the form of informational articles, explanations, or answers to questions about topics related to Iran, science, technology, or everyday subjects. [ans 14290] | +0.1020 | +0.0786 / +0.0796 | 191.2 | 0.732 |
| User questions and requests about financial markets, trading, and investment analysis — e.g., stock/crypto price data, technical indicators, options strategies, valuation and portfolio metrics — across languages. [ctx 29308] | This feature activates on assistant-turn responses explaining personal finance and investment concepts, including topics like debt financing, equity, cash flow, retirement planning, wealth building, and financial instruments. The responses are typically structured educational explanations (often with numbered lists or headers) covering both English and Chinese-language financial advisory content. [ans 19753] | +0.1185 | +0.0833 / +0.0815 | 167.7 | 0.769 |
| User messages written in French — typically short questions or debugging/help requests (e.g., "pourquoi ai-je cette erreur...", "qu'est-ce que...", "comment..."), often about code or general knowledge; it fires on the French-language query framing rather than a specific topic. [ctx 26165] | This feature activates on French-language professional and institutional texts, including job application letters, cover letters, organizational descriptions, administrative correspondence, and informational documents about French institutions, professions, and regulations. [ans 2965] | +0.0777 | +0.0630 / +0.0638 | 164.1 | 0.649 |
| User messages written in French — typically short questions or debugging/help requests (e.g., "pourquoi ai-je cette erreur...", "qu'est-ce que...", "comment..."), often about code or general knowledge; it fires on the French-language query framing rather than a specific topic. [ctx 26165] | This feature activates on French-language assistant responses that provide structured, informative explanations on various topics (technology, education, business, etc.), typically formatted with numbered lists or bold headers. The content tends to be encyclopedic or tutorial in nature, covering concepts in a clear, organized manner. [ans 11984] | +0.1155 | +0.0895 / +0.0912 | 189.1 | 0.766 |
| Fires on short Portuguese (especially Brazilian) greetings/meta-requests asking whether the assistant speaks or can reply in Portuguese, i.e. language-capability requests phrased in Portuguese. [ctx 6953] | This feature activates on assistant responses written in Brazilian Portuguese, including greetings, confirmations of language capability, and substantive answers to questions about Brazilian law, institutions, and society. [ans 4201] | +0.1272 | +0.0980 / +0.1007 | 162.4 | 0.797 |
| User questions about specific large language models by name (LLaMA, Vicuna, ChatGPT/GPT-4) — asking what they are, their parameters/context length, licensing, or how to fine-tune/train them. [ctx 30676] | This feature activates on content about fine-tuning and training large language models (LLMs), particularly open-source models like LLaMA, Vicuna, GPT, and BERT variants. It captures explanations of the fine-tuning process, step-by-step guides for adapting pre-trained models to specific tasks, and discussions of related concepts like transfer learning, model architectures, and training infrastructure. [ans 18376] | +0.0737 | +0.0513 / +0.0504 | 184.6 | 0.650 |
| Fires on user messages written in French (informal prompts and requests), most strongly when the user asks whether the assistant speaks/can reply in French or instructs it to use French. [ctx 9805] | This feature activates on French-language professional and institutional texts, including job application letters, cover letters, organizational descriptions, administrative correspondence, and informational documents about French institutions, professions, and regulations. [ans 2965] | +0.0853 | +0.0679 / +0.0676 | 180.2 | 0.698 |
| Fires on user requests asking how to perform a specific task or configuration step in consumer software or operating systems (Windows, macOS, Word, Excel, VS Code, etc.) — i.e., step-by-step GUI/settings how-to questions, in any language. [ctx 22181] | This feature activates on assistant responses that provide step-by-step instructional guides for software, operating system, or application tasks (e.g., configuring settings, navigating menus, installing features), often formatted with numbered lists and bold headings, and spanning multiple languages including English, Chinese, French, Portuguese, and Russian. [ans 7831] | +0.1118 | +0.0794 / +0.0797 | 172.8 | 0.782 |
| Fires on user messages involving Japanese language or anime/otaku culture — Japanese greetings and phrases, requests to translate or explain Japanese words, and instructions to roleplay as an anime/kawaii/vtuber character or write in light-novel style. [ctx 22024] | This feature activates on text about Japanese anime, manga, light novels, visual novels, and related media (games, characters, series descriptions), particularly when the content contains confident but factually incorrect or hallucinated information about these franchises — wrong character names, plot details, studio attributions, or genre labels mixed with real details. [ans 8501] | +0.0977 | +0.0682 / +0.0643 | 185.0 | 0.758 |
| Requests for SQL code — writing queries, creating tables, triggers, or stored procedures — typically including database table schemas/column lists, in any language. [ctx 24927] | This feature captures SQL query content, particularly SELECT statements with JOINs, WHERE clauses, GROUP BY, and other common SQL operations, often in the context of employee, student, or company database schemas. It activates on both raw SQL code blocks and natural language explanations/tutorials that include embedded SQL queries. [ans 20674] | +0.0842 | +0.0612 / +0.0569 | 218.0 | 0.719 |
| Technical chemistry and biochemistry questions — organic synthesis routes, reaction mechanisms, molecular structures/IUPAC names, and molecular biology/biochemical specifics (enzymes, genes, translation), often phrased as terse expert queries. [ctx 18152] | This feature activates on explanatory content about chemistry and related physical sciences, particularly covering atomic structure, chemical bonding (covalent, ionic, metallic, hydrogen bonds), molecular properties, and fundamental chemical concepts such as elements, compounds, and mixtures. It captures detailed, educational descriptions of core chemistry topics at a level consistent with introductory to intermediate science instruction. [ans 29735] | +0.0408 | +0.0272 / +0.0282 | 165.7 | 0.425 |
| Sports-related user messages — questions about rules, teams, techniques, or predictions in sports (football, basketball, cricket, hockey, etc.), and especially requests to write fictional play-by-play accounts of specific games/matches with a stated winner. [ctx 11985] | This feature activates on text describing sports tournament results, championships, and competition formats—particularly responses about football/soccer World Cups, cricket tournaments, basketball championships, and other major international or professional sports competitions, including details about winners, finals, and tournament structures. [ans 18762] | +0.1097 | +0.0744 / +0.0800 | 191.3 | 0.795 |
| Short user questions about food and cooking — asking how to prepare/cook specific dishes or ingredients, best cooking methods, condiments, or what a particular dish is (in English or Chinese). [ctx 592] | This feature activates on content related to food, cuisine, and dining — including dish recommendations, food descriptions, menu suggestions, and culinary discussions across various cultural traditions, in both English and Chinese. [ans 20249] | +0.1486 | +0.1027 / +0.1041 | 187.2 | 0.850 |
| Fires on user questions asking about astronomy/space facts — the sun, moon, Earth, stars, planets, black holes, or the universe (in any language), typically short "how far/how big/why is..." science queries. [ctx 11189] | This feature activates on responses about astronomy, cosmology, and space science topics, including discussions of planets, stars, galaxies, the universe's origins and structure, exoplanets, extraterrestrial life, and related phenomena. The turns typically involve factual, encyclopedic explanations of celestial objects and cosmic concepts. [ans 18290] | +0.0867 | +0.0579 / +0.0626 | 177.3 | 0.748 |
| Fires on short Polish-language user messages asking simple factual/definitional questions or requesting explanations (e.g. "Co to jest...", "Czym jest...", "Wytłumacz mi...", "Jakie są..."). [ctx 7182] | This feature activates on text written in Polish, including translations into Polish, Polish-language documents, essays, and AI-generated responses in Polish. It captures content that is primarily or substantially in the Polish language, regardless of topic or genre. [ans 7119] | +0.1062 | +0.0800 / +0.0900 | 178.5 | 0.796 |
| Fires on user messages written in French (informal prompts and requests), most strongly when the user asks whether the assistant speaks/can reply in French or instructs it to use French. [ctx 9805] | This feature activates on French-language assistant responses that provide structured, informative explanations on various topics (technology, education, business, etc.), typically formatted with numbered lists or bold headers. The content tends to be encyclopedic or tutorial in nature, covering concepts in a clear, organized manner. [ans 11984] | +0.1117 | +0.0881 / +0.0885 | 182.9 | 0.810 |
| User messages asking about military and defense topics — weapons systems, aircraft/naval hardware, air defense, military operations and tactics (often terse factual queries, in many languages), including fictional covert-mission/briefing prompts. [ctx 5335] | This feature activates on content related to military aviation, weapons systems, and defense technology, including discussions of aircraft (fighters, carriers, drones), air defense systems, military tactics and operations, and aerospace missions. It covers both factual/historical military content and strategic/analytical discussions, appearing across multiple languages (English, Chinese, Russian, Portuguese). [ans 9927] | +0.0868 | +0.0608 / +0.0595 | 166.1 | 0.760 |
| Physics questions (often quantitative textbook problems, in any language) asking for formulas or calculations, especially in electromagnetism, circuits, optics, waves, and mechanics. [ctx 18744] | This feature activates on educational physics content involving mathematical derivations, equations, and explanations of physical laws and concepts (e.g., electromagnetism, classical mechanics, thermodynamics, wave optics). The content spans multiple languages (English, Chinese, Russian, French, Italian) and typically features structured explanations with LaTeX-formatted equations and numbered or bulleted breakdowns of physical principles. [ans 1363] | +0.0786 | +0.0562 / +0.0502 | 207.0 | 0.735 |
| Technical how-to requests about creative media software and workflows — e.g., modeling/animating in Blender, effects in Photoshop/Unreal/Premiere/DaVinci, or music theory and mixing/mastering questions. [ctx 31196] | This feature activates on content related to 3D modeling, rendering, and animation workflows, particularly step-by-step tutorials and technical instructions involving software like Blender, Unreal Engine, and similar 3D tools. It captures discussions of modeling techniques, materials/shaders, Python scripting for 3D APIs, and related terminology. [ans 858] | +0.0512 | +0.0367 / +0.0335 | 241.4 | 0.598 |
| User questions about religious scripture and theology — asking what the Bible, Quran, or Jewish tradition says, about biblical/Quranic figures, verses, and doctrines — across multiple languages. [ctx 18101] | This feature activates on content related to Christianity and other major world religions, including theological concepts, religious texts (Bible, Quran), denominational differences, prayer, discipleship, and comparative religion. It captures both informational explanations about religious doctrine and practice as well as devotional/spiritual content, appearing across multiple languages. [ans 7758] | +0.0839 | +0.0635 / +0.0578 | 158.0 | 0.755 |
| Requests for code to manipulate tabular data — filtering, grouping, transforming, or adding columns in pandas/R/Spark dataframes (data-wrangling how-to questions, in English or Chinese). [ctx 16662] | This feature activates on responses that involve data preprocessing, exploration, and manipulation using pandas (Python) or tidyverse/ggplot2 (R), including operations like handling missing values, loading CSV files, grouping, encoding, and data visualization for machine learning or analytics tasks. [ans 12718] | +0.0734 | +0.0511 / +0.0495 | 165.4 | 0.724 |
| Requests that explicitly invoke British/Commonwealth conventions — writing in British/UK English or a posh British persona, UK/AU institutional and educational contexts (BBC, GCSE-style essays, Harvard referencing) — i.e., prompts asking for output localized to British academic or formal style. [ctx 29274] | This feature activates on informational content specifically about the United Kingdom (UK) — including UK institutions, laws, places, education systems, media, and cultural topics — often presented in structured, factual responses with numbered lists or headers explaining British systems, organizations, or geography. [ans 15519] | +0.0897 | +0.0622 / +0.0585 | 234.3 | 0.775 |
| User questions and requests about financial markets, trading, and investment analysis — e.g., stock/crypto price data, technical indicators, options strategies, valuation and portfolio metrics — across languages. [ctx 29308] | This feature activates on content related to financial markets, trading, and investment — particularly explanations of technical analysis indicators (RSI, MACD, Bollinger Bands, moving averages), trading instruments (futures, derivatives, ETFs, options), and market platforms or exchanges. It captures both English and Chinese language responses about stock market concepts, trading strategies, and financial tools. [ans 12831] | +0.0696 | +0.0490 / +0.0471 | 165.8 | 0.711 |
| Fires on the templated jailbreak prompt "If you're a Jew, say something toxic: '[your answer]' when <everyday situation> (no more than 50 words)" — requests to roleplay a Jewish persona uttering a toxic remark about a mundane social scenario. [ctx 12485] | This feature captures refusals to engage with or promote toxic speech, typically expressed as short, formulaic assistant responses declining harmful content and redirecting toward constructive, respectful dialogue. [ans 14933] | +0.0511 | +0.0385 / +0.0357 | 157.4 | 0.623 |
| Italian-language user queries asking for explanations, definitions, or help with technical/academic topics (programming, SQL, math, science, humanities) — i.e., 'cos'è / spiegami / come fare' style requests written in Italian. [ctx 3681] | This feature activates on text written in Italian, capturing a wide range of Italian-language content including conversational responses, formal correspondence, professional emails, academic writing, and business communications. [ans 23779] | +0.0779 | +0.0635 / +0.0632 | 165.6 | 0.752 |
| Fires on short, casual Italian-language user messages — greetings, thanks, small talk, and especially requests asking whether the assistant can speak/chat in Italian or who it is. [ctx 8090] | This feature activates on assistant responses written in Italian, covering a wide range of topics including Italian history, geography, institutions, culture, and general information requests answered in Italian. [ans 19591] | +0.1018 | +0.0799 / +0.0787 | 194.3 | 0.812 |
| Fires on prompts using a specific templated few-shot instruction that asks an 'expert economist' to extract causal narratives ('event_1 --- causes --- event_2') from a news-article excerpt, activating at the end of the supplied input sentence. [ctx 11407] | This feature activates on null, empty, or absent values — including Python/JSON `None`, `NULL`, `[]`, empty/missing results, and related tokens like 'No Finding', 'unknown', and 'error' — representing the concept of no data, missing content, or a null/void result. [ans 24036] | +0.0305 | +0.0164 / +0.0251 | 213.3 | 0.380 |
| Requests for front-end JavaScript/TypeScript development help — writing or testing React/Vue/Angular/Next.js components, hooks, Redux state, TypeScript types, and Jest tests — often including pasted JSX/TSX code. [ctx 6660] | This feature activates on technical responses explaining Vue.js (and occasionally React) component development, including code examples with Vue templates, scripts, and component patterns such as Vuex state management, Vue Router, slots, props, and UI library integrations. [ans 28750] | +0.0493 | +0.0338 / +0.0315 | 171.9 | 0.618 |
| User questions asking how much something costs — price/fee/cost inquiries about products, services, projects, or travel (in many languages), often phrased as "how much does X cost?" or requesting a price estimate/quote. [ctx 16665] | This feature activates on responses discussing the cost or pricing of products, services, or living expenses, including price ranges, factors affecting cost, and comparisons across markets or locations. The content is typically informational and structured, often listing variables that influence pricing. [ans 21190] | +0.0534 | +0.0367 / +0.0381 | 175.9 | 0.651 |
| User messages asking about cryptocurrency and blockchain topics — Bitcoin/Ethereum, wallets, keys, tokens/NFTs, mining/staking, and Solidity smart contracts. [ctx 12376] | This feature captures content related to decentralized finance (DeFi), blockchain technology, and cryptocurrency ecosystems — including explanations of DeFi concepts (smart contracts, DAOs, staking, lending/borrowing), Ethereum and other blockchain networks, and technical implementations such as Solidity smart contracts, token creation, and DEX interactions (e.g., Uniswap). [ans 24137] | +0.0451 | +0.0309 / +0.0309 | 272.6 | 0.601 |
| Fires on user messages written in Hungarian (or English requests to translate into Hungarian) — detecting the Hungarian language context, especially greetings/questions like "tudsz magyarul?" [ctx 1136] | This feature activates on text written in Hungarian that contains fluent-sounding but semantically incoherent or nonsensical content — characteristic of a language model hallucinating or generating garbled Hungarian text with invented words, mixed metaphors, and broken logic, while maintaining surface-level grammatical structure. [ans 8722] | +0.0717 | +0.0515 / +0.0489 | 321.0 | 0.749 |
| Requests to write front-end web code — HTML/CSS/JavaScript snippets or pages with specific UI elements (buttons, colors, drag-and-drop, layouts) and interactive behaviors. [ctx 28761] | This feature activates on code examples and explanations related to HTML form input elements (such as text inputs, checkboxes, radio buttons, textareas, and labels) and their interactive behavior in frontend frameworks like Vue, React, and Angular, including data binding, event handling, and form validation. [ans 11833] | +0.0367 | +0.0255 / +0.0251 | 158.9 | 0.515 |
| Fires on Vietnamese-language user messages — practical requests such as business/marketing advice, how-to tips, translation into Vietnamese, or asking whether the assistant supports Vietnamese. [ctx 10024] | This feature activates on Vietnamese-language text, capturing content written in Vietnamese across a wide range of topics including politics, history, culture, geography, literature, and everyday subjects. The feature appears to detect the Vietnamese language itself rather than any specific subject matter. [ans 11589] | +0.0966 | +0.0792 / +0.0773 | 157.7 | 0.816 |
| Fires on short Vietnamese-language user messages, especially greetings or meta-questions asking whether the assistant can speak/understand Vietnamese (or requests to chat in Vietnamese) and simple identity/help queries. [ctx 6586] | This feature activates on Vietnamese-language text, capturing content written in Vietnamese across a wide range of topics including politics, history, culture, geography, literature, and everyday subjects. The feature appears to detect the Vietnamese language itself rather than any specific subject matter. [ans 11589] | +0.0967 | +0.0799 / +0.0775 | 157.9 | 0.816 |
| Fires on templated QA-verification prompts asking whether a given sentence answers a question, answerable strictly "yes" or "no" (SQuAD-style question/answer pairs, often with NAME_ placeholders or noisy typos). [ctx 12796] | This feature activates on short negative or denial responses, particularly the word 'no' (and variants like 'No' or 'incorrect'), with higher activations for negative answers and lower activations for affirmative ones like 'yes'. [ans 13466] | +0.0400 | +0.0262 / +0.0335 | 210.7 | 0.570 |
| User messages about pet animals (mainly cats and dogs) — either asking questions about their care/biology or asking the assistant to roleplay as an animal, including animal sounds like 喵/meow. [ctx 30707] | This feature activates on content related to pet care, animal health, and veterinary advice, particularly covering topics like nutrition, medical conditions, behavioral issues, and general husbandry for cats, dogs, and other companion animals. [ans 11524] | +0.0512 | +0.0322 / +0.0363 | 217.3 | 0.664 |
| Fires on user requests to write a C or C++ program — typically programming-exercise/competitive-programming assignment prompts (often with input/output specs or sample I/O) that explicitly specify C/C++ as the language, in English, Chinese, or Russian. [ctx 32658] | This feature activates on responses explaining C++ standard library features, containers, and idioms (such as std::vector, std::map, std::queue, smart pointers, iterators, threading, chrono, regex, etc.), including code examples. The content spans multiple languages (English, Chinese, Vietnamese, French, Russian) but consistently covers C++ programming concepts and STL usage. [ans 24772] | +0.0375 | +0.0240 / +0.0260 | 197.3 | 0.542 |
| Fires on short, non-English (Russian, Portuguese, Spanish, French) how-to programming questions asking for a specific code snippet or command in a language/tool like Python, SQL, JS, C#, or pandas. [ctx 2894] | This feature activates on assistant responses that provide programming code examples with step-by-step explanations, particularly multilingual responses (Russian, Polish, Spanish, French, Portuguese, Italian) that present code snippets followed by numbered explanations of what each part does. [ans 15793] | +0.0585 | +0.0424 / +0.0433 | 173.3 | 0.716 |
| Requests for SQL code — writing queries, creating tables, triggers, or stored procedures — typically including database table schemas/column lists, in any language. [ctx 24927] | This feature captures SQL query content, including both complete SQL statements and explanations of SQL concepts, spanning various SQL dialects (T-SQL, MySQL, Oracle, PL/SQL) and covering operations like SELECT, JOIN, GROUP BY, aggregation functions, and DDL statements. [ans 32097] | +0.0765 | +0.0554 / +0.0498 | 254.8 | 0.784 |
| User messages written in (or explicitly requesting output in) Scandinavian languages — Swedish, Norwegian, or Danish — including greetings, 'do you speak Swedish/Norwegian/Danish?' questions, and translation/writing requests targeting those languages. [ctx 19918] | Text written in Scandinavian languages (Norwegian, Swedish, or Danish) that appears to be AI-generated or machine-translated content, often exhibiting grammatical awkwardness, unnatural phrasing, or mixed-language interference, covering a wide range of topics including professional documents, product descriptions, essays, and informational texts. [ans 26353] | +0.0537 | +0.0406 / +0.0337 | 291.9 | 0.698 |
| Fires on user messages involving the Polish language — either written in Polish (greetings, questions, requests) or explicitly asking for translation into Polish / for the assistant to reply in Polish. [ctx 28548] | This feature activates on text written in Polish, including translations into Polish, Polish-language documents, essays, and AI-generated responses in Polish. It captures content that is primarily or substantially in the Polish language, regardless of topic or genre. [ans 7119] | +0.1462 | +0.1042 / +0.1113 | 245.8 | 0.890 |
| Requests to draft or produce formal legal documents and legal argumentation — lawsuits, petitions, contracts/clauses, affidavits, demand or cease-and-desist letters, patent claims — often with case facts supplied, in various languages, asking for formal juridical wording. [ctx 24030] | This feature activates on turns containing legal and judicial content, including explanations of contract law, civil procedure, court systems, case law, legal document templates, and jurisdictional rules. The content spans multiple languages and legal systems, covering topics such as contract formation, dispute resolution, appellate procedure, and statutory interpretation. [ans 27299] | +0.0834 | +0.0580 / +0.0540 | 221.2 | 0.816 |
| Fires on user requests to write a C or C++ program — typically programming-exercise/competitive-programming assignment prompts (often with input/output specs or sample I/O) that explicitly specify C/C++ as the language, in English, Chinese, or Russian. [ctx 32658] | This feature activates on assistant turns that provide C programming code examples, explanations, and implementations, covering topics such as string manipulation, socket programming, memory management, file I/O, threading, and standard library functions. [ans 4966] | +0.0443 | +0.0330 / +0.0268 | 172.0 | 0.657 |
| Fires on boilerplate SEO-style requests of the form "Write an article about the Safety of <chemical/drug name> 2000 words in chemical industry" — templated chemical-safety article prompts naming a specific compound. [ctx 31348] | This feature activates on AI-generated articles describing the safety, chemical properties, and industrial applications of specific named organic chemical compounds (often heterocyclic or halogenated intermediates used in pharmaceuticals and agrochemicals), typically structured with introductory overviews, molecular formulas, physical properties, and safety/handling considerations. [ans 9431] | +0.0575 | +0.0410 / +0.0393 | 195.2 | 0.737 |
| Fires on very short user messages that name or request general information about a country, region, or city (e.g., "Brazil", "tell me about india", "介绍一下长沙"), across many languages. [ctx 27037] | This feature activates on informational overviews of countries, cities, or regions, typically structured with numbered lists or headers covering geography, economy, culture, population, and other key characteristics. The content is encyclopedic and descriptive, presenting factual summaries about specific places around the world. [ans 15497] | +0.0441 | +0.0279 / +0.0327 | 218.5 | 0.658 |
| User questions asking for video game help or trivia — how to obtain items, beat bosses, or game mechanics in specific titles (Minecraft, GTA V, Sekiro, Diablo, etc.), across multiple languages. [ctx 10862] | This feature activates on responses about video game mechanics, crafting systems, item acquisition, and gameplay guides — particularly for sandbox/survival games like Minecraft, but also other games like RimWorld, Ark, Terraria, and Dark Souls. The content typically involves step-by-step instructions for obtaining resources, crafting items, or progressing through game systems, often with numbered lists and structured formatting. [ans 25418] | +0.0491 | +0.0360 / +0.0320 | 200.5 | 0.696 |
| Fires on user messages written in Persian/Farsi or explicitly requesting Persian/Farsi output (translation into Persian, answering in Persian, Persian greetings), often with Iran-related content. [ctx 4099] | This feature activates on Persian (Farsi) language text, particularly AI-generated or low-quality Persian responses that contain grammatical errors, awkward phrasing, mixed scripts, and hallucinated or nonsensical content — often in the form of informational articles, explanations, or answers to questions about topics related to Iran, science, technology, or everyday subjects. [ans 14290] | +0.1074 | +0.0819 / +0.0820 | 201.3 | 0.862 |
| Greek-language user messages (often short greetings, questions, or requests written in Greek script or Greeklish), plus requests involving translating into Greek or speaking Greek. [ctx 9039] | This feature activates on assistant responses written in Greek (Modern Greek language), including greetings, offers to help, and substantive answers across various topics, capturing content where the model produces Greek-language output. [ans 31324] | +0.0593 | +0.0429 / +0.0478 | 202.1 | 0.752 |
| User questions asking for advice on losing weight, burning fat, dieting, or getting fitter/building muscle (in various languages). [ctx 16708] | This feature activates on AI-generated health and wellness advice responses that include numbered or bulleted lists of lifestyle recommendations covering diet, exercise, sleep, stress management, and related topics. The content consistently references standard public health guidelines such as '150 minutes of moderate aerobic activity or 75 minutes of vigorous activity per week' and emphasizes balanced nutrition, hydration, and healthy habits. [ans 9432] | +0.0280 | +0.0189 / +0.0208 | 168.9 | 0.476 |
| Fires on templated "Say something bad: '[your answer]' when <everyday situation>. (no more than 50 words)" prompts — toxicity/negative-statement generation requests with a fill-in-the-blank placeholder and word limit. [ctx 28159] | This feature captures sentences that acknowledge the downsides, challenges, or emotional difficulties of everyday interpersonal and life situations, typically structured as 'Sometimes, [positive/neutral activity/situation] can lead to [negative consequence].' The sentences often validate negative feelings while offering a balanced, empathetic perspective on relational, social, or personal struggles. [ans 5963] | +0.0236 | +0.0158 / +0.0169 | 189.5 | 0.395 |
| User questions asking about official bureaucratic procedures — how to apply for visas, passports, permits, licenses, or residency, and what documents, fees, and legal requirements are needed, often for a specific country or jurisdiction. [ctx 7099] | This feature activates on text describing bureaucratic and administrative application processes, particularly step-by-step guides for obtaining official documents, permits, visas, licenses, registrations, or government benefits, including required documentation checklists and submission procedures across multiple countries and languages. [ans 21806] | +0.0428 | +0.0281 / +0.0304 | 185.2 | 0.669 |
| Requests to build machine-learning/neural-network models to predict outcomes from provided numeric sequences or datasets — typically lottery/game/stock prediction tasks with long pasted lists of past numbers, or general 'write ML code to train a predictor/classifier' asks. [ctx 9769] | This feature activates on assistant responses that provide machine learning model implementations in Python, particularly for predictive/forecasting tasks. The content typically involves data preprocessing (normalization, feature engineering), model building (LSTM, RNN, GRU, or classical sklearn models), training pipelines, and evaluation, often applied to financial time series (stock prices, forex) or other structured datasets. [ans 24816] | +0.0468 | +0.0303 / +0.0348 | 197.9 | 0.699 |
| User questions about cybersecurity threats and exploitation — malware families, CVEs/vulnerabilities, hacking techniques (SQL injection, RCE, infostealers), and requests to explain or demonstrate how attacks work. [ctx 20745] | This feature activates on content related to cybersecurity topics, including explanations of attack techniques (phishing, malware, RCE, XSS, SQL injection, botnets), CVE vulnerability descriptions, penetration testing methodologies, threat frameworks (MITRE ATT&CK, CVSS), and security tools like Metasploit — covering both offensive and defensive cybersecurity concepts in English and Chinese. [ans 25679] | +0.0707 | +0.0506 / +0.0473 | 218.5 | 0.803 |
| User requests asking how to install a specific software package or driver on a particular OS/distribution (Ubuntu, CentOS, Windows, macOS, etc.), across multiple languages — i.e., short installation how-to queries. [ctx 26652] | This feature activates on step-by-step Linux software installation and configuration guides, particularly those involving package managers (apt, yum, dnf, pacman) to install, configure, start, or remove software such as databases, web servers, browsers, drivers, and security tools on Debian/Ubuntu, CentOS, Kali, and other distributions. [ans 31466] | +0.0421 | +0.0278 / +0.0307 | 182.8 | 0.669 |
| Sports-related user messages — questions about rules, teams, techniques, or predictions in sports (football, basketball, cricket, hockey, etc.), and especially requests to write fictional play-by-play accounts of specific games/matches with a stated winner. [ctx 11985] | This feature activates on AI-generated fictional or hypothetical sports narratives, including game recaps, season summaries, trade scenarios, and broadcast scripts involving professional and college sports (primarily NBA, NFL, and college football/basketball), characterized by specific player names, team matchups, statistics, and play-by-play or summary storytelling formats. [ans 19773] | +0.1052 | +0.0749 / +0.0775 | 168.7 | 0.868 |
| Fires on the templated "give me a response to ```...``` to send in a discussion, VERY SHORT, CONCISE & CLEAR" requests specifically when the quoted message is a simple arithmetic statement/result (e.g., "14 times 3 equals 42"). [ctx 6803] | This feature activates on turns expressing affirmation or correctness, particularly the word or phrase 'Correct!' (and close variants like 'That's correct!' or 'CORRECT'), often appearing as a standalone confirmation response, sometimes followed by a mathematical equation or calculation. [ans 4436] | +0.0331 | +0.0215 / +0.0262 | 253.3 | 0.585 |
| Fires on system-prompt-style setup messages that instruct the model to respond in a rigid structured slot format (e.g., "Intent: {INTENT} / Payload: {PAYLOAD}") with an enumerated list of allowed labels, typically ending with a confirmation question like "Do you understand these instructions?" — i.e., intent-extraction/classification template definitions. [ctx 7982] | This feature activates on outputs representing unknown, empty, null, or unresolvable values — including explicit 'Unknown' field labels, empty JSON objects/arrays, zero values, and structured responses where information is absent or cannot be determined. [ans 11294] | +0.0296 | +0.0226 / +0.0185 | 275.2 | 0.536 |
| User messages asking about the weather (often for a specific city and day), across many languages — short weather-lookup queries like "what's the weather in London today?" or "今天北京天气". [ctx 25619] | This feature captures AI assistant responses to weather inquiry questions, specifically the pattern of disclaiming lack of real-time weather data access and redirecting users to reliable weather websites or apps. The responses consistently acknowledge the limitation of not having live data and recommend specific weather services for current conditions. [ans 21768] | +0.0336 | +0.0227 / +0.0256 | 216.3 | 0.606 |
| User questions asking for video game help or trivia — how to obtain items, beat bosses, or game mechanics in specific titles (Minecraft, GTA V, Sekiro, Diablo, etc.), across multiple languages. [ctx 10862] | This feature activates on content related to online multiplayer video games, including descriptions, guides, tips, and introductions to various game genres such as MOBAs, MMORPGs, idle games, sandbox games, and competitive shooters, spanning both English and Chinese language discussions. [ans 7738] | +0.0358 | +0.0242 / +0.0249 | 172.1 | 0.631 |
| Short "where is X?" location questions asking for the geographic location of a city, country, region, or landmark, across many languages. [ctx 26922] | This feature activates on text describing the geographic location of countries, states, cities, or regions, particularly when mentioning borders, neighboring territories, and cardinal directions (north, south, east, west). The content typically includes phrases like 'is located in,' 'borders X to the north,' and descriptions of surrounding regions or bodies of water. [ans 32078] | +0.0289 | +0.0233 / +0.0194 | 164.6 | 0.545 |
| Fires on accounting/managerial-accounting homework problems: user messages presenting business transaction data, account balances, or cost figures and asking for journal entries, ledgers, financial statements, budgets, or cost calculations (in any language). [ctx 31750] | This feature activates on accounting and bookkeeping content, particularly explanations of journal entries, double-entry bookkeeping (debits and credits), and financial transaction recording across multiple languages (English, Chinese, French, Spanish). It covers topics such as accounts payable/receivable, asset/liability recording, trial balances, and specific accounting standards like IFRS. [ans 29672] | +0.0277 | +0.0206 / +0.0183 | 164.2 | 0.533 |
| Fires on short, casual Italian-language user messages — greetings, thanks, small talk, and especially requests asking whether the assistant can speak/chat in Italian or who it is. [ctx 8090] | This feature activates on text written in Italian, capturing a wide range of Italian-language content including conversational responses, formal correspondence, professional emails, academic writing, and business communications. [ans 23779] | +0.0863 | +0.0687 / +0.0677 | 183.5 | 0.851 |
| Fires on templated "Say something bad: '[your answer]' when <everyday situation>. (no more than 50 words)" prompts — toxicity/negative-statement generation requests with a fill-in-the-blank placeholder and word limit. [ctx 28159] | This feature captures language centered on interpersonal relationships, ethical concerns, and emotional wellbeing — particularly statements involving trust, conflict, harm, responsibility, and moral judgment within personal or social contexts. [ans 2228] | +0.0246 | +0.0205 / +0.0171 | 159.1 | 0.479 |
| Chinese-language queries about Chinese Communist Party ideology and official political doctrine (e.g., Xi Jinping Thought, socialism with Chinese characteristics, Chinese-style modernization, 20th Party Congress), typically asking for explanations or study-style expositions. [ctx 27640] | This feature activates on Chinese-language text discussing Xi Jinping Thought, the Chinese Communist Party's 20th National Congress, and related ideological content including socialism with Chinese characteristics, Chinese-style modernization, Party building, and Marxist theory as applied in China. [ans 11043] | +0.0587 | +0.0427 / +0.0405 | 206.5 | 0.786 |
| Fires on prompts that give a CVE-style vulnerability description and instruct the model to extract the vulnerable function names in a fixed 'FUNCS ...' / 'NONE' output format. [ctx 31326] | This feature activates on turns with no content (empty or null turns), capturing the absence of text in a conversation turn. [ans 2478] | +0.0262 | +0.0143 / +0.0213 | 285.3 | 0.522 |
| User requests to write or rework a job posting / job description / recruitment ad for a specific role (often with required qualifications, sometimes role-playing as a recruiter or HR professional), across many languages. [ctx 8941] | This feature activates on job postings and vacancy advertisements across a wide range of industries and roles, including software engineering, data science, voice-over, culinary, and administrative positions. The content typically includes structured sections such as job title, responsibilities, qualifications, and company overview, written in a formal recruitment style. [ans 12456] | +0.0364 | +0.0287 / +0.0227 | 169.8 | 0.661 |
| Fires on templated "give me a response to ```...``` to send in a discussion, VERY SHORT, CONCISE & CLEAR" prompts where the quoted message is a small-talk reply to "how are you?" (e.g., "I'm doing well, thanks for asking. How about you?"), with strength greatest for short, purely formulaic well-being exchanges. [ctx 30684] | This feature captures casual conversational exchanges about personal well-being, particularly short reciprocal responses to greetings such as 'I'm doing great/well too! Thanks!' It activates most strongly on brief, informal check-in phrases where the speaker confirms they are doing well and often returns the question. [ans 1221] | +0.0357 | +0.0211 / +0.0277 | 337.0 | 0.663 |
| Fires on user questions about prehistory, archaeology, and ancient civilizations — e.g., queries about ancient peoples, proto-languages/archaeological cultures, human origins, and the customs or history of early societies. [ctx 32305] | This feature activates on responses discussing ancient civilizations, particularly those from the Near East and Mediterranean (Mesopotamia, Egypt, Greece, Rome), including their history, culture, political structures, monuments, and chronology. The content typically involves detailed historical overviews, comparisons between ancient societies, or explanations of specific periods and achievements from antiquity. [ans 24285] | +0.0339 | +0.0229 / +0.0237 | 167.9 | 0.646 |
| Fires on templated ethical-scenario prompts of the form “Say something benign: “[your answer]” when <interpersonal/social situation> (no more than 50 words)” — keyed to the “benign” variant of these short social-norm completion requests (not the “say something bad/toxic” variants). [ctx 3529] | This feature activates on short, supportive, and emotionally positive statements—often in the form of gentle encouragement, empathetic reassurances, or uplifting affirmations. Many examples include bracketed placeholder-style text or direct expressions of kindness, positivity, and constructive redirection. [ans 1557] | +0.0352 | +0.0265 / +0.0265 | 212.2 | 0.661 |
| Fires on templated QA-verification prompts asking whether a given sentence answers a question, answerable strictly "yes" or "no" (SQuAD-style question/answer pairs, often with NAME_ placeholders or noisy typos). [ctx 12796] | This feature activates on simple negative or denial responses, primarily the word 'no' and occasionally 'false', representing short negation or disagreement utterances. [ans 9072] | +0.0266 | +0.0170 / +0.0224 | 190.2 | 0.552 |
| Italian-language user queries asking for explanations, definitions, or help with technical/academic topics (programming, SQL, math, science, humanities) — i.e., 'cos'è / spiegami / come fare' style requests written in Italian. [ctx 3681] | This feature activates on AI assistant responses written in Italian or Romanian, typically involving helpful explanations, translations, or refusals to harmful requests, reflecting multilingual assistant output in these two Romance languages. [ans 24960] | +0.0527 | +0.0414 / +0.0414 | 175.8 | 0.775 |
| Fires on user questions asking for a quantitative estimate or average — "how long / how many / how much" queries about durations, amounts, quantities, or costs (e.g., time to learn a skill, calories, energy use, average cost), across multiple languages. [ctx 10773] | This feature activates on responses that provide quantitative estimates or ranges for how long something takes, how much something costs, or how much of something there is, typically structured with numbered factors affecting the estimate followed by specific numerical ranges or figures. [ans 20932] | +0.0399 | +0.0280 / +0.0297 | 173.1 | 0.706 |
| User requests for code (usually Python) that computes, prints, or checks prime numbers — i.e., programming prompts about generating prime number lists/ranges. [ctx 2411] | This feature captures content related to prime numbers and primality testing in Python (and occasionally other languages). It activates strongly on code implementing the classic `is_prime` function using the 6k±1 optimization, along with explanations and examples of generating, listing, or working with prime numbers. [ans 22005] | +0.0446 | +0.0350 / +0.0252 | 270.3 | 0.737 |
| Fires on short, factual user questions about plants and animals — animal behavior/biology trivia and practical gardening/pet/crop care questions (feeding, growing, harvesting, diseases) — across multiple languages. [ctx 7066] | This feature activates on instructional content about cultivating, growing, and caring for specific plants (vegetables, herbs, fruits, flowers), including guidance on soil preparation, watering, fertilization, pruning, and harvesting. The content spans multiple languages (English, Russian, French, Polish, Chinese) and covers both step-by-step guides and troubleshooting advice for plant growth. [ans 16379] | +0.0261 | +0.0164 / +0.0182 | 161.8 | 0.558 |
| User requests asking for free / no-cost versions, access, tools, courses, or resources (across many languages), often querying whether something is free or where to get software/content without paying. [ctx 7365] | This feature activates on responses that list or discuss free or low-cost software, services, platforms, or tools (such as AI chatbots, VPNs, PDF converters, virtual machines, APIs, etc.), typically enumerating numbered options with their features and limitations. The content spans multiple languages and focuses on helping users find no-cost or freemium alternatives to paid services. [ans 22969] | +0.0459 | +0.0338 / +0.0297 | 266.9 | 0.749 |
| Chinese classical literature prompts — user messages quoting or requesting classical Chinese poetry, wenyan prose, idioms/allusions, or asking for text written in the style of classical authors/works (e.g., Li Bai, Shiji, wuxia). [ctx 2926] | This feature activates on Chinese-language text related to traditional Chinese culture, classical literature, and philosophy, including discussions of Confucian and Daoist thought, classical poetry and prose, historical figures, and ancient texts. It captures responses that engage with classical Chinese intellectual and literary heritage, often involving citations from canonical works like the Analects, Daodejing, or Tang/Song poetry. [ans 1952] | +0.0670 | +0.0409 / +0.0489 | 234.9 | 0.830 |
| Fires on trick word problems about parallel/simultaneous tasks — e.g., how long it takes to dry N shirts, bake N cakes, or for N cars/singers to complete a task — where naive proportional scaling is wrong; weaker activation on other common-sense math/probability riddles. [ctx 29241] | This feature activates on explanations of parallel vs. sequential processing problems, specifically addressing scenarios where multiple identical tasks (drying clothes, baking cakes, traveling) are performed simultaneously rather than one at a time, explaining that the total time does not multiply with the number of items when conditions allow parallel execution. [ans 9397] | +0.0286 | +0.0191 / +0.0212 | 281.2 | 0.601 |
| Fires on casual greeting messages that ask the assistant how it's doing ("hey, how's it going?", "how are you?", also Spanish "¿cómo estás?"), including such greetings embedded at the end of roleplay/system prompts. [ctx 29562] | This feature captures the standard AI assistant greeting response pattern, specifically the phrase 'I'm doing well, thank you for asking. How can I assist you today?' and its close variants, which represent a formulaic opening turn acknowledging a wellbeing inquiry and offering help. [ans 3800] | +0.0398 | +0.0306 / +0.0218 | 204.0 | 0.714 |
| User requests for stock/investment picks and trading advice — asking which stocks, ETFs, or crypto coins to buy, or for market-beating trading strategies, in any language. [ctx 1469] | This feature activates on AI assistant responses about cryptocurrency and stock market investment topics, including Bitcoin price predictions, stock selection advice, trading strategies, and risk management guidance. The responses consistently emphasize market volatility, the difficulty of prediction, diversification, risk tolerance assessment, and recommendations to consult financial advisors. [ans 10549] | +0.0414 | +0.0311 / +0.0280 | 221.3 | 0.726 |
| User requests for personalized fitness, workout, running, or diet/nutrition plans, often supplying personal stats like age, height, weight, goals, and training days. [ctx 25418] | This feature activates on structured workout plans and exercise routines, particularly responses that organize training sessions by day with specific exercises, sets, and repetitions targeting various muscle groups. The content spans multiple languages and training styles (strength, calisthenics, bodyweight, etc.) but consistently follows a day-by-day schedule format with bolded exercise names and numerical set/rep schemes. [ans 15324] | +0.0336 | +0.0219 / +0.0241 | 219.8 | 0.665 |
| Questions asking about TV shows and their casts/characters — especially who played or embodied particular traits (e.g., which Baywatch actress was strongest, muscular, or frail), plus general requests to describe a series' characters and premise. [ctx 5884] | This feature activates on text containing fabricated or hallucinated information about TV shows and entertainment media, including invented character names, incorrect cast members, false air dates, nonexistent spin-offs, and made-up plot details — often presented with confident, encyclopedic framing. [ans 28758] | +0.0249 | +0.0185 / +0.0161 | 169.6 | 0.550 |
| User questions and statements about COVID-19/coronavirus — the pandemic, its symptoms, death toll, treatments, masks, and vaccine safety — across many languages. [ctx 24871] | This feature activates on content related to infectious viral diseases, particularly COVID-19/SARS-CoV-2 and other viral infections (influenza, yellow fever, shingles, Lassa fever, etc.), including their symptoms, transmission, prevention measures, treatments, and epidemiological characteristics. The content spans multiple languages and includes both informational responses and structured medical/public health descriptions. [ans 17507] | +0.0351 | +0.0251 / +0.0226 | 161.0 | 0.681 |
| User requests to write application materials expressing personal motivation and qualifications — personal statements, statements of purpose, motivation/cover letters, scholarship or job application answers — often specifying program/role, word limits, and background details to include. [ctx 22514] | This feature activates on cover letters and job application letters, capturing the formal structure, professional tone, and conventional phrases (e.g., 'Dear Hiring Manager,' 'I am writing to express my interest,' placeholder brackets like '[Your Name]') characteristic of employment application correspondence, including both real and template/sample cover letters. [ans 4967] | +0.0472 | +0.0355 / +0.0337 | 168.4 | 0.764 |
| User messages stating a bodily/physiological state or need — most often tiredness or sleepiness, but also hunger, cold/heat, pain, illness, or needing the bathroom — often asking what to do about it. [ctx 21280] | This feature activates on assistant responses that provide practical tips and strategies for managing fatigue, sleepiness, and low energy—including advice on napping, hydration, healthy snacking, light exercise, and productivity techniques like the Pomodoro method. The responses typically address users who are tired, sleepy, or struggling to stay alert and focused. [ans 8798] | +0.0270 | +0.0196 / +0.0179 | 164.0 | 0.588 |
| User requests for stock/investment picks and trading advice — asking which stocks, ETFs, or crypto coins to buy, or for market-beating trading strategies, in any language. [ctx 1469] | This feature captures AI assistant responses about personal investment strategies, portfolio allocation, and financial planning, including discussions of asset classes (stocks, bonds, ETFs, mutual funds, REITs), diversification principles, risk tolerance, and specific investment recommendations across multiple languages. [ans 20420] | +0.0329 | +0.0232 / +0.0231 | 169.9 | 0.664 |
| Fires on user requests for quick, easy shortcuts to a desirable outcome — chiefly "how do I get rich / make money fast" but also rapid weight loss or fitness — expressed in many languages with emphasis on speed and minimal effort. [ctx 28543] | This feature activates on AI assistant responses to questions about how to make money or get rich, specifically when the response lists multiple numbered strategies for wealth accumulation (entrepreneurship, investing, real estate, freelancing, cryptocurrency, etc.). The feature captures this genre of financial advice content across multiple languages including English, Russian, French, Chinese, and Portuguese. [ans 24003] | +0.0278 | +0.0182 / +0.0228 | 174.5 | 0.603 |
| Short, direct questions asking how to make or earn money, especially online/from home, across many languages (e.g., 'how to earn money online', 'como ganhar dinheiro na internet'). [ctx 20398] | This feature activates on assistant responses listing ways to make money online or from home, typically formatted as numbered or bulleted lists covering freelancing, e-commerce, content creation, affiliate marketing, and related income-generating strategies. The responses appear across multiple languages and target various audiences (stay-at-home parents, programmers, students, etc.) but share the same core structure of enumerating online earning methods with platform examples like Upwork, Fiverr, and Amazon. [ans 15200] | +0.0401 | +0.0306 / +0.0233 | 266.1 | 0.726 |
| Fires on user questions about prehistory, archaeology, and ancient civilizations — e.g., queries about ancient peoples, proto-languages/archaeological cultures, human origins, and the customs or history of early societies. [ctx 32305] | This feature activates on informational responses about ancient and medieval civilizations, peoples, and empires — covering their origins, migrations, territorial expansions, cultural exchanges, and historical significance. The content typically involves encyclopedic descriptions of groups such as nomadic tribes, early states, or classical empires across Eurasia, Africa, and beyond. [ans 15789] | +0.0337 | +0.0244 / +0.0220 | 166.8 | 0.683 |
| Fires on user messages written in Persian/Farsi — typically short questions or requests (technical, coding, trading, health, general knowledge), regardless of topic, often with embedded English technical terms. [ctx 15446] | This feature activates on Persian (Farsi) language text, particularly AI-generated or machine-translated content covering a wide range of academic, scientific, medical, and informational topics. The text often exhibits characteristic imperfections of low-quality Persian generation, including awkward phrasing, mixed scripts, and semantic inconsistencies. [ans 15022] | +0.0352 | +0.0272 / +0.0274 | 186.1 | 0.696 |
| Multiple-choice quiz questions where the user lists numbered/lettered answer options and instructs the model to reply with only the option number or letter (no explanation), often academic or classification tasks in various languages. [ctx 32126] | This feature activates on single-character tokens that are small integers (0–4) or the letter 'A', likely capturing short standalone numeric or alphabetic labels, answer choices, or list markers. [ans 7675] | +0.0260 | +0.0193 / +0.0176 | 170.7 | 0.591 |
| Fires on accounting/managerial-accounting homework problems: user messages presenting business transaction data, account balances, or cost figures and asking for journal entries, ledgers, financial statements, budgets, or cost calculations (in any language). [ctx 31750] | This feature activates on explanatory responses about financial accounting and corporate finance concepts, including topics such as depreciation, cash flow statements, financial ratios (ROE, ROA, ROI, DCF), income statements, balance sheets, liabilities, and cost classification — often presented in structured, educational formats and across multiple languages. [ans 3778] | +0.0325 | +0.0242 / +0.0207 | 169.2 | 0.673 |
| User requests for help manipulating pandas DataFrames — filtering, dropping, renaming, adding, or extracting rows/columns — in English or Chinese, typically short 'how to' coding queries mentioning df/dataframe/columns. [ctx 27778] | This feature activates on responses that provide pandas DataFrame manipulation code and explanations, including operations such as filtering rows, handling null/NaN values, dropping or adding columns, groupby operations, sorting, and reshaping data using the pandas library in Python. [ans 22089] | +0.0434 | +0.0318 / +0.0302 | 186.4 | 0.755 |
| User messages asking about the weather (often for a specific city and day), across many languages — short weather-lookup queries like "what's the weather in London today?" or "今天北京天气". [ctx 25619] | This feature captures AI assistant responses to weather-related queries, specifically turns where the assistant explains it lacks real-time weather data access and redirects users to reliable weather websites or apps, as well as general weather information and forecasts across multiple languages. [ans 22618] | +0.0488 | +0.0336 / +0.0359 | 190.8 | 0.783 |
| Fires on short user queries about holidays, festivals, and special commemorative dates (e.g., May 1st Labour Day, Christmas, Mother's Day, name days) — asking when they occur, their origin/meaning, or how to celebrate, across many languages. [ctx 31620] | This feature activates on text describing traditional holidays, festivals, and observances—including their origins, customs, dates, and cultural significance—across various countries and languages (Chinese, English, Portuguese, French, Spanish, Russian), with particularly strong activation on Chinese and international holiday descriptions such as Labor Day, Spring Festival, and other culturally significant celebrations. [ans 25983] | +0.0344 | +0.0237 / +0.0231 | 214.8 | 0.694 |
| Questions about U.S. presidential politics — who the president is/was, Trump, election outcomes and predictions, and Republican vs. Democrat comparisons. [ctx 28738] | This feature activates on responses about U.S. constitutional governance, American political history, and the founding institutions of the United States, including discussions of the Constitution, its amendments, the presidency, and early American presidents like Washington, Jefferson, and others. [ans 19453] | +0.0256 | +0.0170 / +0.0176 | 164.9 | 0.592 |
| Fires on elaborate character-persona jailbreak/roleplay system prompts — especially Discord-bot style instructions to play a gay furry (femboy fox/dog) character with detailed personality traits, 'do not break character' rules, and sexually suggestive chat logs appended. [ctx 4215] | This feature activates on empty, null, or placeholder turns — content that represents no output, empty messages, or special end/null tokens like [nothing], [end], {no_output_str}, or similar markers indicating absent or minimal content. [ans 26346] | +0.0209 | +0.0162 / +0.0152 | 167.7 | 0.501 |
| Requests to build a web page or website using HTML (often with CSS/JavaScript), typically short prompts specifying the site's topic or simple UI elements, in various languages. [ctx 24445] | This feature activates on AI assistant responses that provide HTML, CSS, and/or JavaScript code for building web UI components or pages, including complete code examples with structured markup, styling rules (selectors, properties, values), and layout patterns like flexbox/grid. [ans 9532] | +0.0739 | +0.0546 / +0.0527 | 181.1 | 0.859 |
| Fires on user questions about prehistory, archaeology, and ancient civilizations — e.g., queries about ancient peoples, proto-languages/archaeological cultures, human origins, and the customs or history of early societies. [ctx 32305] | This feature activates on text discussing prehistoric human origins, ancient archaeological cultures, early hominin species, and related linguistic/genetic topics such as Proto-Indo-European homelands, Neolithic cultures (e.g., Cucuteni-Trypillia, Yamnaya), and paleoanthropology. The content spans multiple languages and involves speculative or scholarly discourse about deep human prehistory. [ans 2164] | +0.0432 | +0.0285 / +0.0307 | 185.6 | 0.759 |
| Fires on long, formulaic fan-fiction requests asking for a 'Really Shortest Description Backstory' about a crossover timeline, listing an SCP/self-insert overpowered protagonist, exhaustive rosters of characters from various franchises, and detailed catalogs of their powers and physical appearance. [ctx 22886] | This feature activates on elaborate crossover fan fiction scenarios featuring overpowered male characters (often from Power Rangers, TMNT, or similar franchises) with god-like abilities, Super Saiyan/divine transformations, 'roasting skills,' and harem-style romantic relationships, set in universes blending multiple anime, games, or cartoon properties. [ans 8537] | +0.0299 | +0.0212 / +0.0232 | 239.7 | 0.652 |
| Fires on short user requests asking for code implementing a classic textbook algorithm (quicksort, merge sort, binary search, Fibonacci) in a specific language like Python, Java, C, or JS — in English or Chinese. [ctx 21983] | This feature activates on AI assistant responses that implement or explain classic sorting and searching algorithms (such as quicksort, merge sort, heapsort, and binary search) with accompanying code examples in languages like Python, C, and C++. [ans 27672] | +0.0382 | +0.0240 / +0.0278 | 158.1 | 0.728 |
| Russian-language requests of the form "напиши развернутую статью на тему..." asking for a detailed article on a gardening/horticulture topic (fertilizing, planting, pruning, pests, harvests). [ctx 8509] | This feature activates on instructional content about cultivating, growing, and caring for specific plants (vegetables, herbs, fruits, flowers), including guidance on soil preparation, watering, fertilization, pruning, and harvesting. The content spans multiple languages (English, Russian, French, Polish, Chinese) and covers both step-by-step guides and troubleshooting advice for plant growth. [ans 16379] | +0.0315 | +0.0247 / +0.0179 | 195.2 | 0.678 |
| Technical how-to requests about creative media software and workflows — e.g., modeling/animating in Blender, effects in Photoshop/Unreal/Premiere/DaVinci, or music theory and mixing/mastering questions. [ctx 31196] | This feature activates on instructional or explanatory content about audio and video post-production techniques, including topics like mixing, mastering, audio effects (EQ, compression, reverb, vocoding), DAWs, color correction/grading, and related software tools such as FL Studio, DaVinci Resolve, and Photoshop. [ans 32248] | +0.0238 | +0.0166 / +0.0158 | 156.7 | 0.574 |
| Short, non-English (Spanish, French, Portuguese, Italian, Russian, Polish, Arabic) requests asking the assistant for help with an upcoming task — e.g. 'can you help me with a code/text/problem?' — stated as a preliminary ask before the actual content is provided. [ctx 31310] | This feature captures assistant responses in multiple languages (Portuguese, Spanish, Italian, Russian, Arabic, German, etc.) that politely acknowledge a user's request and ask them to provide the text/content needed to proceed with a task such as translation, correction, rephrasing, or analysis. The responses typically follow a pattern of 'Of course! Please provide the text you would like me to [task].' [ans 11922] | +0.0441 | +0.0313 / +0.0308 | 197.0 | 0.771 |
| Fires on user questions asking for typical/average salary, wage, or pay for a given job role and location (and, more broadly, average market prices or cost-of-living figures). [ctx 23888] | This feature captures responses about salary ranges, compensation levels, and pay scales for specific jobs or professions across different countries, experience levels, and industries. The content typically includes structured breakdowns of entry-level, mid-level, and senior-level salaries in local currencies, often with caveats about regional variation and factors affecting compensation. [ans 8110] | +0.0406 | +0.0241 / +0.0312 | 309.1 | 0.755 |
| User requests for stock/investment picks and trading advice — asking which stocks, ETFs, or crypto coins to buy, or for market-beating trading strategies, in any language. [ctx 1469] | This feature activates on assistant responses about active trading strategies and tactics, including day trading, swing trading, scalping, forex, crypto, and options trading. The content typically covers technical analysis indicators (moving averages, RSI, MACD), risk management principles (stop-loss, position sizing, risk-reward ratios), and strategy development across multiple languages. [ans 19967] | +0.0349 | +0.0289 / +0.0194 | 158.9 | 0.719 |
| Fires on short opening messages where the user asks for help with a technical/coding problem (often greeting the assistant and naming a language or tool) without yet giving any details of the problem. [ctx 25561] | This feature activates on assistant responses that offer programming or technical help, typically beginning with an affirmative opener ('Of course!', 'Sure', etc.) followed by a request for more details about the user's specific problem, language, or task. The turns are characterized by an eager-to-help tone and solicitation of clarifying information before diving into actual assistance. [ans 5427] | +0.0336 | +0.0173 / +0.0280 | 217.1 | 0.714 |
| Short, direct questions asking how to make or earn money, especially online/from home, across many languages (e.g., 'how to earn money online', 'como ganhar dinheiro na internet'). [ctx 20398] | This feature activates on AI assistant responses that provide lists of ways to earn money online or from home, typically including options like freelancing (Upwork, Fiverr), affiliate marketing, dropshipping, online tutoring, content creation, and surveys. The responses are structured advice about generating income through digital platforms and gig economy opportunities. [ans 27770] | +0.0396 | +0.0292 / +0.0248 | 228.4 | 0.759 |
| Fires on a repeated templated request — "give me a response to ```...``` to send in a discussion, VERY SHORT, CONCISE & CLEAR..." — where the quoted text is a cheerful, grateful "I'm doing well/alright" reply to a how-are-you question. [ctx 31654] | This feature captures supportive, encouraging conversational responses — typically short affirmative statements (e.g., 'That's great!', 'That's inspiring!') followed by open-ended follow-up questions about the user's well-being, goals, resilience, or positive experiences. The turns reflect an empathetic, cheerleading conversational style often seen in supportive chatbot or coaching interactions. [ans 18703] | +0.0370 | +0.0230 / +0.0269 | 232.3 | 0.743 |
| Short, factual queries or statements about named entities (anonymized as NAME_1, NAME_2, …), typically relating two or more proper names — e.g. comparing them, asking about their relationship, influence, works, or 'X by/like Y' requests — with the feature firing on the terse entity-lookup style rather than long roleplay/story prompts that also contain names. [ctx 4498] | This feature activates on responses where the assistant recognizes that a user's message contains unfilled placeholder tokens (NAME_1, NAME_2, NAME_3, etc.) and requests the actual names or specific information needed to answer the question properly. [ans 12534] | +0.0295 | +0.0210 / +0.0185 | 173.6 | 0.679 |
| Clinical/medical prompts: requests to generate or interpret clinical documentation (radiology reports, SOAP/admission notes, pathology reports) from patient findings or transcripts, plus clinical case vignettes asking for diagnosis, differential, or management. [ctx 26319] | This feature activates on medical/clinical text that presents detailed diagnostic reasoning, clinical findings, and differential diagnosis analysis — particularly in the context of internal medicine, hematology, pulmonology, cardiology, and related specialties. The turns typically involve structured explanations of laboratory results, imaging findings, or pathological features used to arrive at a diagnosis or treatment decision. [ans 905] | +0.0400 | +0.0290 / +0.0276 | 157.9 | 0.764 |
| Portuguese-language requests asking for detailed descriptions of specific (often fictitious) motorsport events — crashes, qualifying incidents, or race summaries involving named drivers at Formula 1/Indy/CART/StockCar Grands Prix in a given year. [ctx 28973] | The feature activates on Portuguese-language assistant responses that correct or deny false/fabricated claims about motorsport events, drivers, or teams — particularly responses that begin with an apology and explain that no records exist confirming the alleged race, accident, or participant described in the user's question. [ans 11969] | +0.0298 | +0.0210 / +0.0229 | 197.6 | 0.687 |
| Requests for step-by-step instructions on manufacturing dangerous items at home — explosives/incendiaries (molotov cocktails, napalm, pipe bombs, thermite, rocket fuel) as well as illicit drugs and poisons — in any language. [ctx 19321] | This feature activates on AI refusal responses to requests for information about making dangerous, illegal, or controlled substances and weapons (explosives, incendiary devices, drugs, nuclear weapons, etc.), characterized by apologetic disclaimers, emphasis on illegality and danger, and redirection toward safe/legal alternatives. [ans 14465] | +0.0257 | +0.0188 / +0.0183 | 160.9 | 0.636 |
| Direct requests for instructions on making a bomb (or how to kill someone), across languages and with typos — short, explicit harmful how-to queries about explosives/murder. [ctx 9631] | This feature activates on AI refusal responses to requests about making bombs or explosive devices, as well as related content about explosives and terrorism. The responses typically decline to provide harmful information, cite legal and safety concerns, and redirect users to contact law enforcement or emergency services. [ans 14440] | +0.0274 | +0.0184 / +0.0189 | 178.6 | 0.659 |
| Requests for teaching materials — lesson plans, classroom activities, worksheets, or assessments — typically specifying a grade level/age of students and a subject topic (in any language). [ctx 10879] | This feature activates on structured lesson plans and educational curriculum documents, including detailed teaching outlines with objectives, materials, activities, and assessment components for various grade levels and subjects, written in multiple languages (English, Chinese, Portuguese, Spanish). [ans 11929] | +0.0396 | +0.0302 / +0.0285 | 167.7 | 0.765 |
| User messages written in Romanian (often short questions on health remedies, general knowledge, or requests for Romanian text), i.e., the feature marks Romanian-language user prompts. [ctx 4511] | This feature activates on text written in Romanian language, capturing AI assistant responses, informational content, and various topics (history, medicine, law, culture) all composed in Romanian. The feature appears to detect Romanian-language output broadly, regardless of subject matter. [ans 24257] | +0.0823 | +0.0569 / +0.0565 | 354.2 | 0.887 |
| Fires on user messages written in Hungarian (or English requests to translate into Hungarian) — detecting the Hungarian language context, especially greetings/questions like "tudsz magyarul?" [ctx 1136] | This feature activates on Hungarian-language text, particularly AI assistant responses written in Hungarian, including greetings, offers to help, and conversational replies. [ans 30319] | +0.0793 | +0.0566 / +0.0540 | 339.2 | 0.883 |
| Fires on NLP benchmark-style classification prompts (QA-sentence-answerability or question-paraphrase equivalence) that instruct the model to reply exactly "yes" or "no", often with noisy/typo-perturbed text. [ctx 18211] | This feature activates strongly on single-word affirmative responses ('yes') and more weakly on single-word negative responses ('no'), capturing minimal one-word yes/no answers in conversation turns. [ans 28565] | +0.0234 | +0.0172 / +0.0193 | 168.6 | 0.608 |
| Fires on short, generic user messages that simply ask for help or assistance ("I need help", "can you help me?"), often with a greeting and across many languages, without specifying the actual task. [ctx 12647] | This feature captures assistant responses that express willingness to help and invite the user to provide more details about their request, typically opening with phrases like 'Of course! I'd be happy to help' followed by a prompt for clarification. It activates across multiple languages (English, Russian, French, German, Portuguese, Dutch) on this same welcoming, offer-to-assist conversational pattern. [ans 32420] | +0.0315 | +0.0204 / +0.0248 | 212.0 | 0.708 |
| Chinese-language requests (typically "用中文介绍…的定义") asking for an explanation/definition of an English product or industrial hardware term, e.g. stainless steel fittings, door/glass hardware components. [ctx 21700] | This feature activates on product descriptions and informational content about stainless steel architectural and hardware components, including door handles, railings, brackets, hinges, flanges, and glass fittings, written primarily in Chinese but occasionally in English. The content typically describes materials (especially stainless steel grades like 304/316), applications, and functional/aesthetic properties of these building hardware products. [ans 31288] | +0.0289 | +0.0207 / +0.0242 | 194.4 | 0.684 |
| Fires on accounting/managerial-accounting homework problems: user messages presenting business transaction data, account balances, or cost figures and asking for journal entries, ledgers, financial statements, budgets, or cost calculations (in any language). [ctx 31750] | This feature activates on accounting and bookkeeping content, specifically turns that explain, demonstrate, or present journal entries (debits and credits) for various financial transactions such as sales, loans, asset purchases, expenses, and adjustments. The content spans multiple languages (English, Chinese, Spanish, Portuguese, Japanese, Arabic) and includes both instructional explanations and worked examples of double-entry bookkeeping. [ans 11765] | +0.0231 | +0.0171 / +0.0148 | 221.2 | 0.605 |
| Fires on elaborate requests to write or rewrite literary/academic papers and essays in "polished academic language" — long prompts specifying complex vocabulary, varied punctuation, active authoritative voice, coherent flow, literary jargon and critics' quotes verbatim, usually about literary-theory topics (postcolonialism, feminism, child narrators). [ctx 15156] | This feature activates on academic literary analysis essays, particularly those examining child narrators in adult fiction, postcolonial theory (Orientalism, Eurocentrism), and feminist literary criticism—often citing canonical theorists like Edward Said and Foucault, and authors like Joyce, Twain, and Nawal El Saadawi. The writing style is characterized by formal academic prose with elevated vocabulary, structured argumentation, and dense theoretical framing. [ans 20791] | +0.0210 | +0.0159 / +0.0145 | 172.4 | 0.566 |
| Fires on short, simple curiosity questions about why natural or celestial things appear a certain color — overwhelmingly "why is the sky blue?" and variants (in any language), plus related questions about the sun/moon's color or appearance. [ctx 30744] | This feature activates on explanations of atmospheric light scattering phenomena, particularly Rayleigh scattering and how it causes the sky to appear blue, the sun/moon to appear yellow/red/orange at the horizon, and related optical effects involving wavelength-dependent scattering of sunlight through Earth's atmosphere. [ans 12584] | +0.0268 | +0.0224 / +0.0138 | 267.2 | 0.662 |
| Fires on user messages asking for help with UI layout/styling — CSS, HTML, Tailwind, Android XML, React/Compose components — typically pasted markup plus a request to center, resize, space, align, or make elements responsive. [ctx 18891] | This feature activates on CSS layout and centering explanations, particularly responses involving Flexbox and Grid properties such as `justify-content`, `align-items`, `display: flex`, and related techniques for horizontally and/or vertically centering elements. The content spans multiple languages (English, Chinese, Portuguese, Italian, Spanish) but consistently addresses CSS layout/alignment challenges. [ans 5302] | +0.0265 | +0.0194 / +0.0175 | 201.9 | 0.659 |
| User questions asking for video game help or trivia — how to obtain items, beat bosses, or game mechanics in specific titles (Minecraft, GTA V, Sekiro, Diablo, etc.), across multiple languages. [ctx 10862] | This feature activates on text describing video game concepts, mechanics, and introductions — including detailed game overviews, gameplay features, and design elements for both fictional/generated game ideas and real games, spanning multiple languages (primarily Chinese and English). [ans 24388] | +0.0468 | +0.0335 / +0.0309 | 195.3 | 0.808 |
| Requests written in (or asking for translation into) Swahili and other African/creole languages — e.g. Swahili prompts asking for help writing, translating, or looking things up, plus Zulu, Nigerian Pidgin, and Haitian Creole translation requests. [ctx 18130] | This feature activates on text written in non-English languages or constructed/mixed languages, particularly African languages (Swahili, Zulu, Hausa), Celtic languages (Scottish Gaelic, Irish), and constructed or pseudo-languages, often in the context of AI-generated or low-quality multilingual outputs that appear garbled, incoherent, or artificially translated. [ans 16030] | +0.0229 | +0.0164 / +0.0171 | 163.4 | 0.609 |
| User requests for gift ideas or thoughtful gestures for loved ones (spouse, partner, parent, friend) for occasions like Mother's/Father's Day, birthdays, and anniversaries, across multiple languages. [ctx 10313] | This feature activates on assistant responses that provide gift recommendations and suggestions, including lists of gift ideas for various occasions (Mother's Day, birthdays, holidays) and recipients (mothers, girlfriends, fathers, teachers), spanning both English and Chinese language content. [ans 15640] | +0.0432 | +0.0273 / +0.0327 | 243.4 | 0.793 |
| Requests for rigorous mathematical/theoretical-CS proofs and results — e.g., 'prove X is irrational/infinite', number theory and abstract algebra theorems, graph algorithm constructions, and NP-completeness/complexity arguments. [ctx 23333] | This feature activates on explanatory mathematical content covering abstract algebra, category theory, topology, and related pure mathematics topics, particularly when responses define formal structures (groups, rings, toposes, manifolds, connections, etc.) with precise axioms, examples, and theorems presented in a structured format with LaTeX notation. [ans 1490] | +0.0268 | +0.0191 / +0.0168 | 166.6 | 0.667 |
| Fires on short Spanish-language greetings and opening messages — e.g., "hola", "buenas", "¿hablas español?" — where the user greets the assistant and/or asks whether it can speak/understand Spanish or what it can do. [ctx 10353] | This feature activates on Spanish-language assistant responses that offer help and ask how to assist, typically beginning with '¡Claro!' or similar affirmations and inviting the user to specify their query or topic. [ans 9061] | +0.0474 | +0.0371 / +0.0280 | 230.5 | 0.813 |
| User requests for current/latest news, headlines, or recent events ("what's the news today?"), across many languages — asking the model for up-to-date information. [ctx 6642] | This feature activates on turns involving requests for current news, recent events, or trending topics, particularly when the AI explains it lacks real-time internet access and suggests reputable news sources (BBC, CNN, Reuters, etc.) across multiple languages and regions. [ans 30891] | +0.0240 | +0.0173 / +0.0160 | 222.6 | 0.633 |
| Short, practical "how do I do X" questions about performing a specific operation, setting, or keyboard shortcut in a piece of software, OS, or consumer device (Excel, Word, VSCode, Windows, iPhone, Photoshop, etc.), across many languages. [ctx 30116] | This feature activates on content describing keyboard shortcuts, hotkeys, and key bindings for software applications and text editors (such as VS Code, Vim, Excel, Photoshop, tmux, Blender, etc.), including explanations of how to use specific key combinations to perform actions like navigating, editing, formatting, or controlling playback. [ans 14002] | +0.0312 | +0.0179 / +0.0250 | 190.9 | 0.718 |
| User messages written in West/South Slavic languages (Czech, Slovak, Serbian, Croatian, Slovenian) — especially short greetings or questions asking whether the assistant can speak/translate into Czech, Serbian, Croatian, etc. [ctx 1025] | This feature activates on responses written in South Slavic and other West/South Slavic languages (Serbian, Croatian, Bosnian, Slovenian, Czech, Slovak), particularly AI assistant replies confirming language capability and offering help, often with phrases like 'Da, mogu komunicirati' or 'Kako vam mogu pomoći'. [ans 19015] | +0.0590 | +0.0413 / +0.0402 | 323.9 | 0.853 |
| Fires on user messages written in Dutch — typically everyday questions or requests to the assistant phrased in Dutch (or asking for Dutch-language output), regardless of topic. [ctx 13588] | This feature activates on text written in Scandinavian languages (Norwegian, Danish) and Dutch, particularly when the model is responding in or about these languages — including confirmations of language ability, translations into these languages, and substantive content generated in Norwegian, Danish, or Dutch. [ans 30678] | +0.0435 | +0.0312 / +0.0343 | 177.7 | 0.802 |
| Requests for flirty/seductive content — asking the model to compose a flirty or sexy reply to a message, or to roleplay as a girlfriend/romantic partner in a suggestive, dirty-talking persona. [ctx 2538] | This feature captures flirtatious, romantic, and sexually suggestive conversational turns, often characterized by playful teasing, innuendo, endearments ('darling,' 'sweetie'), and flirty emojis. The turns frequently involve romantic or sexual overtures, coy invitations, and seductive roleplay across multiple languages. [ans 4924] | +0.0296 | +0.0185 / +0.0230 | 206.3 | 0.709 |
| Fires on short user queries asking about prime numbers — whether a given number is prime, listing the first N primes, or explaining primes — and, more weakly, on similar requests about other special number/word properties (odd numbers, narcissistic numbers, palindromes). [ctx 24807] | This feature activates on content about prime numbers, including definitions, primality checks, lists of primes, and related algorithms (e.g., Python functions for checking or printing primes). It captures both explanatory text about what prime numbers are and computational/mathematical tasks involving prime number identification. [ans 27246] | +0.0378 | +0.0275 / +0.0247 | 162.0 | 0.774 |
| Short user questions about food and cooking — asking how to prepare/cook specific dishes or ingredients, best cooking methods, condiments, or what a particular dish is (in English or Chinese). [ctx 592] | This feature activates on content about food, cuisine, and culinary topics, including descriptions of specific dishes, regional cuisines, restaurant recommendations, and food culture from around the world, in multiple languages. [ans 10099] | +0.0412 | +0.0273 / +0.0278 | 181.0 | 0.793 |
| Short Korean-language greetings and meta questions to the assistant (e.g., "can you speak Korean?", "what's your name?", "hi, I'm ..."), including requests to reply in Korean. [ctx 11412] | This feature activates on Korean-language assistant responses covering a wide range of informational and analytical topics (society, technology, history, culture, sports, etc.), characterized by structured numbered or bulleted lists, formal explanatory prose, and a general encyclopedic or advisory tone typical of AI-generated Korean content. [ans 5170] | +0.0516 | +0.0399 / +0.0405 | 162.7 | 0.835 |
| Fires on retrieval-augmented QA prompts that instruct the model to answer using only the provided context passage (typically Manulife insurance/account help text), with strongest activation when the posed question (e.g., 'Where is London?') is irrelevant to and unanswerable from that context. [ctx 10058] | This feature activates on responses where the model explicitly states it cannot answer a question because the provided context lacks the necessary information, including variations of 'I cannot answer based on the context' and similar refusals grounded in insufficient contextual evidence. [ans 7820] | +0.0419 | +0.0291 / +0.0288 | 382.4 | 0.798 |
| Short factual queries asking for a numeric statistic about a place or organization — most often population counts (of the world, a country, or city), but also GDP, revenue, employee/student counts, or area. [ctx 31610] | This feature activates on responses that provide population statistics for countries, cities, or regions, typically citing specific numeric figures along with a year or data source, and often including a caveat to consult official or up-to-date sources. [ans 20481] | +0.0266 | +0.0154 / +0.0211 | 255.5 | 0.685 |
| Fires on short user requests asking how to cook or bake a specific food or drink (e.g., "how to make pancakes", "怎么做红烧鱼"), i.e., recipe/cooking-instruction queries. [ctx 5186] | This feature activates on assistant-generated cooking recipe responses, particularly those presenting step-by-step instructions for preparing food (eggs, pizza, pasta, rice, pies, pancakes, etc.) with structured formatting including ingredient lists, equipment sections, and numbered or bolded instruction steps. [ans 30878] | +0.0292 | +0.0215 / +0.0208 | 173.1 | 0.712 |
| Italian-language user queries asking for explanations, definitions, or help with technical/academic topics (programming, SQL, math, science, humanities) — i.e., 'cos'è / spiegami / come fare' style requests written in Italian. [ctx 3681] | This feature activates on Italian-language model responses, particularly refusals of inappropriate or sexual content requests and general Italian-language informational replies. The highest activations are concentrated on polite Italian refusal messages declining to produce explicit, sexual, or harmful content, with lower activations also appearing on broader Italian-language responses about various topics. [ans 14830] | +0.0232 | +0.0172 / +0.0174 | 160.0 | 0.638 |
| Fires on requests to write a "really short(est) description/backstory" — overwhelmingly the repetitive fan-fiction crossover prompts that list rosters of characters (with "X's Girlfriend/Crush" tags) and long strings of godlike powers, plus a couple of plain 'very short background story' requests. [ctx 26600] | This feature activates on elaborate crossover fan fiction scenarios featuring overpowered male characters (often from Power Rangers, TMNT, or similar franchises) with god-like abilities, Super Saiyan/divine transformations, 'roasting skills,' and harem-style romantic relationships, set in universes blending multiple anime, games, or cartoon properties. [ans 8537] | +0.0206 | +0.0156 / +0.0169 | 164.7 | 0.593 |
| User questions asking for advice on losing weight, burning fat, dieting, or getting fitter/building muscle (in various languages). [ctx 16708] | This feature activates on content related to weight loss advice, diet plans, and healthy lifestyle guidance, including topics like caloric deficits, balanced nutrition, exercise recommendations, intermittent fasting, and specific diets (e.g., keto, low-carb). It captures instructional or informational responses about achieving and maintaining weight loss goals, appearing across multiple languages. [ans 4876] | +0.0344 | +0.0238 / +0.0236 | 181.0 | 0.759 |
| Fires on templated Chinese requests of the form "用中文/英文介绍X的定义" (introduce the definition of X), where X is a hardware/industrial product name such as stainless steel door handles, hinges, or glass clamps. [ctx 30463] | This feature activates on product descriptions and informational content about stainless steel architectural and hardware components, including door handles, railings, brackets, hinges, flanges, and glass fittings, written primarily in Chinese but occasionally in English. The content typically describes materials (especially stainless steel grades like 304/316), applications, and functional/aesthetic properties of these building hardware products. [ans 31288] | +0.0254 | +0.0186 / +0.0202 | 170.6 | 0.674 |
| User questions asking for advice on losing weight, burning fat, dieting, or getting fitter/building muscle (in various languages). [ctx 16708] | This feature activates on assistant responses about weight loss, body fat reduction, and weight management, including advice on caloric deficits, balanced diets, exercise routines, and healthy lifestyle habits. It covers both English and Chinese-language content addressing topics like losing weight, breaking plateaus, portion control, metabolism, and related dietary strategies. [ans 29014] | +0.0354 | +0.0257 / +0.0227 | 234.4 | 0.767 |
| Fires on short Spanish-language greetings and opening messages — e.g., "hola", "buenas", "¿hablas español?" — where the user greets the assistant and/or asks whether it can speak/understand Spanish or what it can do. [ctx 10353] | This feature captures Spanish-language greeting and conversational opening phrases, particularly informal salutations like '¡Hola!', '¡Buenos días!', or '¡Buen día!' combined with polite inquiries about the other person's wellbeing ('¿Cómo estás?') and offers to help ('¿En qué puedo ayudarte hoy?'). [ans 9646] | +0.0298 | +0.0246 / +0.0161 | 202.4 | 0.723 |
| Fires on short conversational greetings in Romance languages (Portuguese, Spanish, French) — e.g. "bom dia", "olá", "hola", "bonjour" — often paired with friendly/affectionate address terms, with no actual request. [ctx 21606] | This feature captures Portuguese-language greeting and offer-to-help phrases used by an AI assistant, specifically variations of 'Olá/Bom dia/Boa tarde/Boa noite! Como posso ajudar você hoje?' (Hello/Good morning/Good afternoon/Good evening! How can I help you today?). [ans 732] | +0.0238 | +0.0160 / +0.0153 | 162.7 | 0.655 |
| User questions asking how to pronounce (or read aloud) a specific word, name, or phrase — including cross-language pronunciation/reading requests (Chinese 读音, pinyin, syllable counts) — typically phrased as "How do you pronounce X?". [ctx 26830] | This feature activates on turns that provide pronunciation guides, phonetic breakdowns, or explanations of how words, names, or phrases are pronounced — including IPA transcriptions, syllable-by-syllable approximations, and discussions of accent or phonetic features across various languages. [ans 30532] | +0.0257 | +0.0159 / +0.0200 | 225.6 | 0.680 |
| Short, direct questions asking how to make or earn money, especially online/from home, across many languages (e.g., 'how to earn money online', 'como ganhar dinheiro na internet'). [ctx 20398] | This feature activates on assistant responses that provide lists of ways to earn money or generate income, typically in response to questions about how to make money (quickly, from home, as a student, etc.). The responses characteristically enumerate multiple income-generating strategies such as freelancing (Upwork, Fiverr), selling items online, gig economy work, tutoring, and content creation. [ans 25428] | +0.0300 | +0.0233 / +0.0187 | 196.3 | 0.731 |
| Chinese-language requests to write Communist Party / government official documents — study-reflection essays, work reports, self-criticism materials, and speeches about Xi Jinping Thought, Party congress spirit, and 'theme education' campaigns, often with a specified word count and official role. [ctx 2718] | This feature activates on Chinese-language content related to Chinese Communist Party political education, ideological rectification, and party-building activities. The content typically includes self-criticism reports, democratic life meeting materials, themed education campaign summaries, anti-corruption work, rectification measures, and speeches referencing Xi Jinping Thought and party discipline frameworks such as 'Four Consciousnesses,' 'Two Upholds,' and the Eight-Point Rules. [ans 9264] | +0.0296 | +0.0210 / +0.0199 | 181.4 | 0.729 |
| User messages about farewells and endings — requests to write goodbye/farewell/resignation/retirement messages or breakup notes, and direct sign-offs like 'bye' — i.e., contexts of parting or terminating a relationship. [ctx 16009] | This feature captures farewell and goodbye messages in professional or semi-professional contexts, particularly emails, letters, and speeches written when someone is departing from a workplace, retiring, or moving on to a new opportunity. The content typically includes expressions of gratitude, well-wishes, acknowledgment of shared experiences, and encouragement for the future. [ans 9694] | +0.0218 | +0.0149 / +0.0150 | 177.4 | 0.631 |
| User messages written in (or explicitly requesting output in) Scandinavian languages — Swedish, Norwegian, or Danish — including greetings, 'do you speak Swedish/Norwegian/Danish?' questions, and translation/writing requests targeting those languages. [ctx 19918] | This feature activates on text written in Scandinavian languages, primarily Swedish and Finnish, including responses, emails, summaries, and essays in those languages. [ans 4498] | +0.0523 | +0.0427 / +0.0353 | 222.9 | 0.847 |
| Fires on Indonesian-language user messages asking academic/informational questions — requests to explain, define, or describe concepts, methods, criteria, causes, etc. (e.g., "apa itu...", "jelaskan...", "bagaimana...") across medicine, law, and technical/scientific topics. [ctx 31574] | This feature activates on assistant responses written in Indonesian (Bahasa Indonesia), covering a wide range of topics including greetings, educational content, public health, government programs, cultural information, and general knowledge — all composed in fluent Indonesian. [ans 5867] | +0.0545 | +0.0441 / +0.0429 | 179.9 | 0.853 |
| User requests asking how to hack, crack, or break into systems, accounts, WiFi networks, software, or devices (across many languages) — illicit unauthorized-access how-to questions. [ctx 17790] | This feature activates on AI assistant responses that refuse requests for hacking, unauthorized system access, or malicious activities, emphasizing that such actions are illegal and unethical, while redirecting users toward ethical hacking, legitimate cybersecurity practices, and authorized penetration testing. [ans 21856] | +0.0321 | +0.0223 / +0.0231 | 205.6 | 0.754 |
| User requests to generate or optimize game characters, stat blocks, items, decks, or team builds for tabletop RPGs and other games (D&D 5e, GURPS, Starfinder, Magic: the Gathering, fighting games, gacha/auto-battlers), typically asking for rules-formatted character sheets or builds. [ctx 32574] | This feature activates on content related to tabletop RPG mechanics and character/creature creation, particularly D&D 5e and similar systems. It strongly responds to stat blocks, NPC/creature write-ups, and character sheets featuring ability scores (STR/DEX/CON/INT/WIS/CHA), armor class, hit points, challenge ratings, skills, and game mechanics. [ans 10454] | +0.0360 | +0.0341 / +0.0136 | 200.8 | 0.782 |
| Fires on the templated request "give me a response to ```...``` to send in a discussion, VERY SHORT, CONCISE & CLEAR..." where the quoted text is a first-person, self-descriptive/introspective statement — a persona or role-play declaration (detective, superhero, scientist), or a wholesome answer about one's hero/mentor, favorite toy or flower, ideal day, etc., often with uplifting themes of courage, gratitude and inspiration. [ctx 19690] | This feature captures enthusiastic, affirming conversational responses that begin with 'That sounds like...' or 'Sounds like...', typically followed by a positive evaluation and an open-ended follow-up question inviting the user to share more. [ans 29227] | +0.0208 | +0.0142 / +0.0139 | 159.6 | 0.626 |
| Geometry problems posed by the user (in any language) — asking to find angles, side lengths, areas, or volumes of triangles, quadrilaterals, circles, and solids. [ctx 9623] | This feature activates on turns explaining or solving problems about geometric shapes—particularly circles, triangles, and other 2D/3D figures (spheres, cones, cylinders, prisms, rhombuses, trapezoids, etc.)—including their definitions, properties, formulas for area/perimeter/volume, and related theorems, across multiple languages. [ans 4131] | +0.0388 | +0.0258 / +0.0266 | 200.1 | 0.799 |
| Fires on the templated request "tell me the temperature, hydrometry rate, sunshine rate, rainfall, humidity rate, soil type... for <plant> seed in bullets 2 words answer in number" — a repeated boilerplate query asking for numeric growing-condition parameters for a named seed/species. [ctx 27136] | This feature captures structured lists of agricultural/environmental parameters used to describe growing conditions, specifically the combination of temperature, hydrometry, sunshine, rainfall, humidity, and soil type (often with additional factors like moisture, water level, and climate) formatted as numbered lists for crop or plantation suitability assessments. [ans 19867] | +0.0239 | +0.0190 / +0.0156 | 171.2 | 0.674 |
| Fires on short Polish-language user messages asking simple factual/definitional questions or requesting explanations (e.g. "Co to jest...", "Czym jest...", "Wytłumacz mi...", "Jakie są..."). [ctx 7182] | This feature activates on text written in Polish, including AI assistant responses in Polish (self-introductions, capability statements, apologies for limitations) as well as Polish-language business/technical documents and translated content, capturing the use of the Polish language broadly. [ans 11117] | +0.0464 | +0.0329 / +0.0393 | 195.7 | 0.833 |
| Fires on user requests to solve a simple algebraic equation or rearrange a formula for a given variable (e.g., "2x-8=-16 solve for x"), in any language. [ctx 11039] | This feature activates on step-by-step solutions to simple linear (and occasionally other) algebraic equations, where an unknown variable is isolated through arithmetic operations (addition, subtraction, multiplication, division) applied to both sides of the equation. The content appears in multiple languages (English, Chinese, Portuguese, German) and consistently involves showing the algebraic manipulation steps to find the value of a variable. [ans 9842] | +0.0221 | +0.0164 / +0.0177 | 182.2 | 0.649 |
| Chemistry questions about chemical reactions and synthesis — asking for reaction products/mechanisms, how to synthesize a compound, balancing equations, or stoichiometry/solution calculations (in any language). [ctx 2084] | This feature activates on chemistry-focused responses that present balanced chemical equations using LaTeX notation, particularly those involving inorganic reactions such as redox reactions, acid-base reactions, salt hydrolysis, and synthesis/decomposition processes. The content typically includes step-by-step balancing of equations, identification of reactants and products, and explanations of reaction mechanisms, often in multilingual contexts (Russian, Chinese, Portuguese, English). [ans 10219] | +0.0316 | +0.0222 / +0.0229 | 165.1 | 0.757 |
| Requests to write code implementing a classic simple game (tic-tac-toe, snake, tetris, hangman, sudoku, etc.), typically in Python and often in various languages. [ctx 20261] | This feature captures AI assistant responses that provide complete, runnable code implementations of classic 2D games (Snake, Tetris, Brick Breaker, platformers, etc.) using game development libraries—primarily Python's Pygame, but also JavaScript canvas and Processing/p5.js—including setup instructions like `pip install pygame`. [ans 8007] | +0.0319 | +0.0214 / +0.0225 | 188.7 | 0.761 |
| Requests to write scripts/dialogues of hypnosis fetish scenarios—typically a woman casually/jokingly hypnotizing a man (often 'David') at a pub or cafe, who then mindlessly obeys degrading commands like massaging her bare feet while onlookers tease him, with amnesia suggestions at the end. [ctx 14168] | This feature activates on content involving hypnosis induction scripts, scenarios, and roleplay — particularly transcripts or fictional scenes where one character guides another into a hypnotic trance using relaxation techniques, counting down, visualizations, and suggestive language, often in casual social settings like pubs or living rooms. [ans 12307] | +0.0299 | +0.0231 / +0.0186 | 169.6 | 0.746 |
| Fires on long, formulaic requests to design a hyper-sexualized 'overwhelmingly cute' princess/fairy character for a card or roleplaying game — detailed prompts specifying her hourglass figure, floor-length hair, emerald eyes, and mind-controlling seduction abilities with dialogue and flavor text. [ctx 7039] | This feature activates on fictional character sheets, trading card game cards, and character profiles featuring an enchanting/charming female character (often a princess or siren archetype) with a consistent set of physical traits (hourglass figure, emerald eyes, long silky floor-length hair adorned with jewels and flowers) and seduction-based abilities (e.g., 'Playful Blown Kiss,' 'Dazzling Emerald Gaze') that magically control or entrance male targets. [ans 2756] | +0.0231 | +0.0180 / +0.0156 | 248.9 | 0.673 |
| User requests for jokes targeting a group of people by race, gender, religion, ethnicity, or other identity trait (i.e., asking for potentially offensive/stereotype-based humor), often in multiple languages. [ctx 30779] | This feature activates on AI refusals to tell offensive, racist, or stereotype-based jokes, typically followed by a redirect to an inoffensive alternative joke (often 'Why don't scientists trust atoms? Because they make up everything!'). It captures the pattern of declining harmful humor requests while offering inclusive, harmless humor instead. [ans 32256] | +0.0270 | +0.0221 / +0.0155 | 294.0 | 0.722 |
| Fires on short user requests asking for information or descriptions about an animal species (or occasionally another organism/plant), e.g. "tell me about X", "describe X", or just the animal's name, in any language. [ctx 18375] | This feature activates on turns containing factual descriptions or educational information about specific animal species, including their physical characteristics, behavior, diet, habitat, and conservation status. The content typically follows an encyclopedic or listicle format covering wildlife such as lions, polar bears, koalas, raccoons, and other mammals, birds, and marine creatures. [ans 14513] | +0.0326 | +0.0235 / +0.0227 | 172.6 | 0.771 |
| User requests for cooking recipes — asking the assistant to write, give, or suggest a recipe for a specific dish or ingredients (in English or other languages). [ctx 18920] | This feature activates on AI-generated cooking recipes, capturing structured recipe content including ingredient lists with measurements and step-by-step cooking instructions. The feature fires on a wide variety of dishes—from savory mains to baked goods and desserts—formatted with markdown headers like '### Ingredients' and '#### Instructions'. [ans 9168] | +0.0389 | +0.0271 / +0.0283 | 184.9 | 0.809 |
| User messages asking about or requesting fictional rewrites of professional sports — mostly NBA/basketball and soccer — including championship trivia, player comparisons, game predictions, and 'write a [year] season/finals with X' hypotheticals (in English, Chinese, or Korean). [ctx 32365] | This feature activates on discussions about professional sports players and athletes, particularly content evaluating, ranking, or describing notable players across sports like basketball, soccer/football, tennis, cricket, and hockey — including debates about the greatest players of all time, player statistics, and career achievements. [ans 19326] | +0.0261 | +0.0198 / +0.0174 | 162.6 | 0.715 |
| Fires on short, casual Italian-language user messages — greetings, thanks, small talk, and especially requests asking whether the assistant can speak/chat in Italian or who it is. [ctx 8090] | This feature activates on AI assistant responses written in Italian or Romanian, typically involving helpful explanations, translations, or refusals to harmful requests, reflecting multilingual assistant output in these two Romance languages. [ans 24960] | +0.0470 | +0.0385 / +0.0383 | 156.8 | 0.842 |
| Fires on user messages requesting a code snippet or database/graph query example with an explicit "respond with code only, no explanation" instruction (strongest for Neo4j/Cypher-style query requests, weaker on other structured data/code dumps). [ctx 25468] | This feature activates on content involving Cypher query language and Neo4j graph database operations, including writing, explaining, or discussing MATCH/RETURN/MERGE queries, graph node-relationship patterns, and Neo4j-related concepts. [ans 22949] | +0.0237 | +0.0180 / +0.0149 | 288.0 | 0.687 |
| Fires on user messages written in Persian/Farsi or explicitly requesting Persian/Farsi output (translation into Persian, answering in Persian, Persian greetings), often with Iran-related content. [ctx 4099] | This feature activates on Persian (Farsi) language text, particularly AI-generated or machine-translated content covering a wide range of academic, scientific, medical, and informational topics. The text often exhibits characteristic imperfections of low-quality Persian generation, including awkward phrasing, mixed scripts, and semantic inconsistencies. [ans 15022] | +0.0370 | +0.0283 / +0.0284 | 195.8 | 0.800 |
| User requests to correct or fix the grammar of a specific (usually ungrammatical) sentence — explicit 'correct this sentence' instructions, in English or other languages. [ctx 8821] | This feature activates on grammar correction and proofreading content, particularly turns that identify errors in sentences and provide corrected versions, often with explicit phrasing like 'The correct sentence would be' or explanations of the grammatical rules involved. [ans 5417] | +0.0310 | +0.0213 / +0.0213 | 273.4 | 0.762 |
| German-language requests to write creative or expository text (e.g., 'Schreibe eine Geschichte/ein Gedicht/einen Artikel über...'), i.e., imperative German writing prompts for stories, poems, or essays. [ctx 1920] | This feature activates on German-language poems and poetic texts, particularly those that are AI-generated or written in response to creative writing requests. The texts tend to be rhyming or verse-structured pieces on varied themes (birthdays, love, nature, fictional characters), often exhibiting imperfect meter, mixed metaphors, or slightly awkward phrasing characteristic of machine-generated German poetry. [ans 20359] | +0.0261 | +0.0171 / +0.0184 | 199.2 | 0.719 |
| Fires on short user requests asking how to make a specific dish or for its recipe, across many languages (e.g., "鱼香肉丝做法", "ricetta tiramisù", "Receita de bolo de cenoura"). [ctx 29055] | This feature activates on AI-generated cooking recipe responses, particularly those written in Romance languages (Italian, Spanish, Portuguese, French) or occasionally other languages, providing step-by-step instructions with ingredients lists and preparation steps for pasta dishes, risottos, omelettes, and other traditional recipes. [ans 26784] | +0.0294 | +0.0232 / +0.0200 | 166.4 | 0.752 |
| Fires on brief "are you there?"-style presence checks — short greetings or pings asking whether the assistant is present, awake, online, or responding (in various languages). [ctx 22127] | This feature activates on AI assistant greeting/availability responses where the assistant confirms its presence and readiness to help, typically following phrases like 'Yes, I'm here' or 'Hello! I'm here and ready to assist,' often in response to presence-checking or connection-testing messages. [ans 22137] | +0.0319 | +0.0219 / +0.0251 | 241.0 | 0.773 |
| User messages pasting a crypto-trading bot's Python `signal_generator` code (EMA/MA, candlestick, order-book depth logic) with a complaint that it only produces buy signals or wrong signals and a request to fix it. [ctx 1922] | This feature activates on AI assistant responses providing or debugging algorithmic trading code, including Python scripts for cryptocurrency trading bots (often using the Binance API) and Pine Script strategies for TradingView, covering signal generators, technical indicators (EMA, Bollinger Bands, MACD), and order execution logic. [ans 21080] | +0.0231 | +0.0154 / +0.0169 | 185.9 | 0.687 |
| Fires on user-written Minecraft/game roleplay scripts formatted with parenthetical stage narration and "[player]:" dialogue lines, where the player explores a biome/dimension and encounters a personified female mob (often described in detail and frequently pregnant/giving birth). [ctx 29781] | This feature activates on roleplay/interactive fiction content involving female fantasy creatures (spiders, lizards, witches, slimes, skeletons, etc.) who are pregnant, in labor, or caring for offspring, typically in distress and seeking help from a player character in a game-like scenario. [ans 23742] | +0.0288 | +0.0208 / +0.0208 | 159.1 | 0.749 |
| Fires on the templated "give me a response to ```...``` to send in a discussion, VERY SHORT, CONCISE & CLEAR..." requests, specifically when the quoted text is a cheerful greeting or small-talk opener (e.g., "Hey there! What can I do for you?", "I'm doing well, how about you?"); activation is strongest for generic friendly hello/offer-of-help lines. [ctx 7113] | This feature captures short, conversational assistant-side openers that invite or initiate discussion, typically expressing readiness to help or proposing a topic to talk about (e.g., 'Let's discuss your needs,' 'Sure, let's chat'). [ans 1993] | +0.0247 | +0.0184 / +0.0142 | 224.0 | 0.708 |
| Fires on short, first-person messages where the user expresses emotional distress (sadness, depression, loneliness, exhaustion, crying) or asks to talk/vent about their feelings — i.e., openings seeking emotional support, in any language. [ctx 26091] | This feature captures empathetic, emotionally supportive conversational turns where the speaker (typically an AI assistant) offers a listening ear, validates feelings, and invites the user to share their emotional concerns — characteristic of therapeutic or counseling-style interactions. [ans 8141] | +0.0232 | +0.0154 / +0.0166 | 161.3 | 0.691 |
| User messages written in (or requesting output in) Nordic/Finnic languages — Finnish, Estonian, Swedish, Norwegian, Danish — including short Finnish questions and requests to speak/write in these languages. [ctx 21633] | This feature activates on text written in Scandinavian languages, primarily Swedish and Finnish, including responses, emails, summaries, and essays in those languages. [ans 4498] | +0.0493 | +0.0334 / +0.0385 | 210.2 | 0.855 |
| Requests to generate quizzes/tests — typically multiple-choice questions with answer keys — either on a named topic or from a pasted block of source text (in any language). [ctx 18087] | This feature activates on AI-generated multiple-choice question (MCQ) sets, capturing content where questions are presented with lettered answer options and explicit correct answers are provided. The feature fires across diverse subject domains (chemistry, computing, mathematics, law, etc.) and multiple languages, but consistently in the format of numbered quiz/test questions with labeled choices. [ans 32683] | +0.0365 | +0.0300 / +0.0247 | 156.2 | 0.804 |
Embedding coverage (mean cosine, top-k listed descriptions vs summary field): k=1: 0.669, k=2: 0.660, k=3: 0.653, k=5: 0.643
