[Goodfire](/content/ "Goodfire"/index.html)

Llama 3.1 8B SFT SAE Feature Viewer×Dolci DPO DataLocal view

LocalGlobal

Dolci DPO Data × Llama 3.1 8B SFT SAEHypothesis Testing & Alignment Baking

—

prompt pairs

—

features

—

feature clusters

—

prompt clusters

20000 prompt pairs (chosen − rejected delta)

color:Δ per prompt clusterprompt clusterΔ for response cluster

— jump to response feature cluster (sorted by max |Δ|) —#616 · |Δ|max=0.464− · Tokens and punctuation that mark structure within code, JSON, XML, and#756 · |Δ|max=0.409− · Small tokens and punctuation in programming and technical code context#207 · |Δ|max=0.400− · Very common short function words, punctuation, and small code/markup t#623 · |Δ|max=0.345− · Small punctuation, placeholders, and markup tokens commonly found insi#812 · |Δ|max=0.328+ · Tiny formatting and punctuation fragments commonly found in academic, #805 · |Δ|max=0.314+ · Generic tokens that commonly appear in encyclopedic, legal, and produc#685 · |Δ|max=0.288− · Code and developer documentation: programming examples, how‑to instruc#173 · |Δ|max=0.281− · End-of-code tokens and small punctuation artifacts that mark code bloc#231 · |Δ|max=0.275− · Tokens and syntax from web and configuration markup — HTML/JSX tags, J#780 · |Δ|max=0.263− · Markers and tokens that indicate structured, technical, or tabular doc#592 · |Δ|max=0.261+ · Very common short words, punctuation, and everyday proper-name fragmen# 65 · |Δ|max=0.259+ · LaTeX-formatted mathematics and technical notation found in scholarly # 50 · |Δ|max=0.247+ · LaTeX-formatted mathematical and scientific expressions common in rese#807 · |Δ|max=0.246− · Short common words and punctuation used as generic connectors in runni#556 · |Δ|max=0.239− · Technical source code, markup, and structured documentation fragments #653 · |Δ|max=0.236− · High-frequency function words, punctuation, and markup tokens common i#776 · |Δ|max=0.236+ · Formal academic and technical writing (scientific, medical, and profes# 18 · |Δ|max=0.235+ · Scholarly and legal writing with formal citations, case references, te#390 · |Δ|max=0.233− · Tokens common in scientific, technical, and academic papers — citation#811 · |Δ|max=0.232+ · Text rich in numbers, mathematical expressions, and technical numeric #789 · |Δ|max=0.224+ · Very common English function words and generic assistant meta-phrases #286 · |Δ|max=0.220− · Bits of programming and markup examples — common code snippets, loop/c#291 · |Δ|max=0.220− · Tokens and short snippets from programming/configuration files — ident#806 · |Δ|max=0.219+ · Finding or selecting numeric coefficients when simplifying or rearrang#559 · |Δ|max=0.219− · Small code tokens and punctuation found inside source code and technic#529 · |Δ|max=0.218+ · Short, high-frequency tokens and identifiers common in programming and#429 · |Δ|max=0.218+ · Short common words, abbreviations, and technical acronyms found across#217 · |Δ|max=0.216− · Programming source-code tokens and common identifier/markup elements (#248 · |Δ|max=0.213+ · Formal LaTeX-heavy academic mathematics and physics writing with displ#446 · |Δ|max=0.211+ · Formatting and rendering tables, tabular data, and table-like section #660 · |Δ|max=0.211+ · Table and tabular numeric entries and headers in scientific/medical pa#729 · |Δ|max=0.210+ · Very common small words and punctuation used to link, refer to, or str#305 · |Δ|max=0.209+ · Table-like formatting and numeric entries in scientific, technical, an# 0 · |Δ|max=0.209− · Very common short words, punctuation, and generic connective tokens fo#783 · |Δ|max=0.209+ · Very common short function words and grammatical tokens used in academ#385 · |Δ|max=0.206− · Structured lists, outlines, and step-by-step instructional content (bu#615 · |Δ|max=0.204+ · Common short function words, punctuation, and connective tokens found #601 · |Δ|max=0.204+ · Formal mathematical and technical academic writing (LaTeX-style equati#452 · |Δ|max=0.200+ · Tokens and short fragments typical of academic and scientific papers (#499 · |Δ|max=0.200+ · LaTeX and mathematical notation found in academic papers (math-mode to#796 · |Δ|max=0.198+ · Short numeric tokens, simple arithmetic expressions, and tiny common p#330 · |Δ|max=0.196+ · Mentions of people's names and name tokens in formal, encyclopedic, ac#795 · |Δ|max=0.195− · Words and phrases that add or introduce additional points in formal/in#404 · |Δ|max=0.194+ · Common words, phrases, and tokens used in formal academic math/physics#482 · |Δ|max=0.194+ · Formal academic and scientific paper prose: LaTeX-style research writi#787 · |Δ|max=0.193+ · Technical code and configuration text: program snippets, logs, numeric#391 · |Δ|max=0.192− · Programming and technical Q&A content with code snippets, configuratio#657 · |Δ|max=0.192+ · Small code/math tokens and index placeholders appearing in technical f#798 · |Δ|max=0.191+ · Small numeric tokens and digit literals (single‑ and double‑digit numb#650 · |Δ|max=0.191+ · Formal scientific, mathematical, and technical writing with equations,#518 · |Δ|max=0.188+ · Technical and programming document content — source code, configuratio#464 · |Δ|max=0.186+ · Discussion of laws, regulations, government policy, and compliance req#652 · |Δ|max=0.186+ · Instructional and explanatory content with headings, lists, and common#251 · |Δ|max=0.186− · Graphic sexual content and erotic scenes, often extreme/taboo (non-con# 14 · |Δ|max=0.184− · Technical and academic writing, especially scientific papers and progr#732 · |Δ|max=0.183+ · Very common short English words and grammatical tokens (function words#437 · |Δ|max=0.183+ · Helpful programming tutorials and troubleshooting: short code examples#558 · |Δ|max=0.183+ · Very common short function words and grammatical tokens appearing in f#700 · |Δ|max=0.182+ · In-text academic citations, reference lists, and author name tokens in#308 · |Δ|max=0.181+ · Academic and technical research writing about mathematical, statistica#423 · |Δ|max=0.180+ · Very frequent short function words and punctuation that appear across #588 · |Δ|max=0.180+ · Instructional step-by-step guidance, lists, and structured how-to cont#630 · |Δ|max=0.179+ · Tokens and phrases common in academic, technical, and code-heavy docum#483 · |Δ|max=0.178− · Common tokens and small symbols in programming examples (identifiers, #629 · |Δ|max=0.175+ · Formal academic and technical writing in mathematics and related STEM #484 · |Δ|max=0.174+ · Text that points readers to external web resources, links, examples, o#778 · |Δ|max=0.173+ · Formal academic and scientific prose — common small words, connectors,#808 · |Δ|max=0.173+ · Formal encyclopedic or biographical text with named entities and commo#536 · |Δ|max=0.173+ · Formal academic and professional writing (scientific, legal, and techn# 35 · |Δ|max=0.173+ · Very common short function words and generic connective tokens found a#118 · |Δ|max=0.173+ · Mentions of URLs, web links, and web-related address fragments in tech#411 · |Δ|max=0.172+ · Topics about healthcare delivery, hospitals, and health policy (care s#531 · |Δ|max=0.171+ · Formal informational and technical writing (academic, scientific, pate#506 · |Δ|max=0.171+ · Narrative scene descriptions focusing on locations, sudden sounds, and# 10 · |Δ|max=0.170+ · Common short function words and tiny connective tokens used in everyda#259 · |Δ|max=0.169+ · Polite assistant introductions, self-descriptions, and safe refusals t#497 · |Δ|max=0.169+ · Helpful, safety-conscious assistant responses giving practical advice,#587 · |Δ|max=0.169− · Short function words, punctuation, and small numeric tokens used in te#635 · |Δ|max=0.167− · Short numeric and alphanumeric tokens used as indices, IDs, counts, or#346 · |Δ|max=0.167+ · Technical source-code and API documentation phrasing—comments, docstri#788 · |Δ|max=0.166− · Technical and academic texts with numbers and inline reference markers#358 · |Δ|max=0.166+ · Technical and academic writing tokens and formatting common in researc#737 · |Δ|max=0.165+ · Small syntactic tokens and identifier patterns inside programming sour#813 · |Δ|max=0.165+ · Tokens and places inside programming source code (identifiers, punctua# 8 · |Δ|max=0.162+ · Formal academic mathematics writing — LaTeX-style research prose with # 45 · |Δ|max=0.160+ · Explicit sexual content, often violent or taboo (non-consensual, inces#649 · |Δ|max=0.160− · Technical and scientific documents with code, tables, symbols, and for#782 · |Δ|max=0.160− · Tokens and identifiers in programming examples — variable/function nam#590 · |Δ|max=0.160− · Very high-frequency small English words and tiny punctuation/formattin#680 · |Δ|max=0.160− · Short punctuation, formatting, and small code/document tokens that mar#641 · |Δ|max=0.159+ · Phrases that use “let” to invite, prompt, transition, or soften statem#712 · |Δ|max=0.159+ · Scientific, technical, and code text that uses numbers and short commo#738 · |Δ|max=0.159+ · Technical and academic text tokens—numbers, labels, code/math symbols,# 91 · |Δ|max=0.158+ · Common tokens and short word fragments (place names, proper nouns, sho#282 · |Δ|max=0.157+ · Fragments and tokens from programming source code: identifiers, keywor#222 · |Δ|max=0.157+ · Words and name fragments that begin with the letters S‑ followed by a #666 · |Δ|max=0.156+ · Words and short phrases used in neutral explanatory or comparative pro#218 · |Δ|max=0.156+ · Very common short words, subword fragments, and other high-frequency t#139 · |Δ|max=0.155+ · Phrases and responses enforcing safety rules—refusing or restricting h#255 · |Δ|max=0.155+ · Code snippets and programming help showing common source-code tokens, #619 · |Δ|max=0.155+ · Very common short words, punctuation, and frequent name fragments that#194 · |Δ|max=0.155+ · Sexual and erotic scenes—often explicit, taboo, or non-consensual cont#734 · |Δ|max=0.154+ · General-purpose English words and tokens common in encyclopedic, scien#699 · |Δ|max=0.154+ · Phrases and wording about limits, restrictions, and stated limitations#772 · |Δ|max=0.154+ · Transforming and simplifying algebraic expressions that combine repeat#684 · |Δ|max=0.154+ · Academic and scientific methods/results wording describing study desig#186 · |Δ|max=0.153+ · Technical academic writing in theoretical and experimental high-energy#797 · |Δ|max=0.153+ · Very common small words and generic programming terms used to link or #800 · |Δ|max=0.153− · Recognizing and parsing dates, numeric fields, and other structured me#298 · |Δ|max=0.152− · Generic high-frequency small words and punctuation used as connective #810 · |Δ|max=0.152+ · Phrases expressing caution, doubt, or reserved expectations about peop#758 · |Δ|max=0.152− · Requests and operations involving symbolic derivatives and simple poly#802 · |Δ|max=0.152+ · Neutral linking language used in scientific or factual descriptions—co#428 · |Δ|max=0.151+ · Practical programming and developer how-to content: installation, code#250 · |Δ|max=0.151+ · Numbered or ordered steps and list markers used in instructions, proce#792 · |Δ|max=0.151+ · Very common short English words and small grammatical/connective token#103 · |Δ|max=0.150+ · Small, common code tokens and syntax patterns found in programming exa#804 · |Δ|max=0.150− · Very common sentence-level tokens and connective words that mark ordin#519 · |Δ|max=0.150− · References to lists and list-handling in technical and programming con#200 · |Δ|max=0.149+ · Very common English function words and generic connective tokens that #294 · |Δ|max=0.149+ · Scientific and clinical research writing reporting measurements, stati#225 · |Δ|max=0.148+ · Formal academic, scientific, and technical writing (research papers, p#697 · |Δ|max=0.148+ · Snippets of programming source code and build/compile-related tokens a#686 · |Δ|max=0.147+ · Common short function words and filler tokens that structure everyday #510 · |Δ|max=0.147+ · Common small function words and generic phrasing that appear across fo# 3 · |Δ|max=0.146+ · General-purpose English words and phrases used in factual, technical, #116 · |Δ|max=0.146+ · Assistant-generated Slavic-language replies and informal conversationa#367 · |Δ|max=0.146+ · Words and names starting with the letters 'L'+'o/Li/Lu/Les' appearing #215 · |Δ|max=0.145+ · Mentions of government, public institutions, and formal official bodie#264 · |Δ|max=0.145+ · Tokens that begin with the letters H–Ha/He/Hi/Ho/Ho* — common names, p#128 · |Δ|max=0.145+ · News and opinion writing about political controversy, public backlash,#110 · |Δ|max=0.145+ · Common function words, connective phrasing, and routine prose found in#413 · |Δ|max=0.145− · Instructions, examples, and structured headings used in technical or h#491 · |Δ|max=0.145+ · Code patterns for conditional checks and guarding against missing/inva#594 · |Δ|max=0.145+ · Very high‑frequency English function words and short grammatical token#593 · |Δ|max=0.144+ · References to geographical place names and proper nouns, especially th#801 · |Δ|max=0.144+ · Narratives depicting coercive, exploitative sexual situations and comm#770 · |Δ|max=0.143+ · Very high-frequency small English words and short functional tokens (c#262 · |Δ|max=0.142+ · Tokens and snippets from programming source code—variable names, type/# 71 · |Δ|max=0.142+ · Technical writing about electronic and optoelectronic hardware, fabric#379 · |Δ|max=0.142− · Short common words, UI labels, and section/headings found in web pages#767 · |Δ|max=0.142+ · Very frequent short English words and punctuation used as generic conn#419 · |Δ|max=0.142+ · Headlines, titles, and section/heading language common to news, academ#645 · |Δ|max=0.141+ · Phrases and short function-word expressions used to introduce, explain# 97 · |Δ|max=0.141+ · Words, names, and short letter sequences starting with the letters G–G#566 · |Δ|max=0.141+ · Common small words and numeric tokens used in factual, technical, and #131 · |Δ|max=0.141+ · Very common short function words and conversational fillers that appea#673 · |Δ|max=0.140+ · Common short function words and everyday connective phrases used in ne#730 · |Δ|max=0.140+ · Phrases and transitions that refer to time—past events, recent finding#353 · |Δ|max=0.140+ · Numerical data and numeric table entries in scientific, medical, techn#448 · |Δ|max=0.140+ · Everyday conversational words and pronouns used in instructions, clari#779 · |Δ|max=0.139+ · Academic and scientific writing (formal papers in math/physics/condens#612 · |Δ|max=0.138− · Practical advice and reminders phrased as cautious, helpful instructio#744 · |Δ|max=0.138+ · Very common short function words and simple discourse connectors (basi# 77 · |Δ|max=0.138+ · Very common short function words and small grammatical tokens used acr#475 · |Δ|max=0.138+ · Explicit sexual content and sexual violence, especially incestuous or #584 · |Δ|max=0.138+ · Mentions of organizations, institutions, and formal roles (companies, #790 · |Δ|max=0.138+ · Language and prompts about steps, step-by-step procedures, and increme#755 · |Δ|max=0.138+ · Very high-frequency small English words and grammatical tokens used in#268 · |Δ|max=0.137+ · Beginner-friendly programming how‑to answers with concrete code exampl# 83 · |Δ|max=0.137+ · Very common short function words and everyday filler language found in#221 · |Δ|max=0.137+ · Phrases where the assistant introduces itself, states limits or safety#326 · |Δ|max=0.137+ · Common code patterns for accessing, assigning, updating, and iterating#399 · |Δ|max=0.137+ · Source code files and developer-facing documentation comments (code ex#799 · |Δ|max=0.137+ · Phrases and words used to express farewells, departures, loss, and say#803 · |Δ|max=0.136+ · Text about historical, political, and medical topics using common smal#505 · |Δ|max=0.136+ · Common short function words and negation tokens used to qualify, limit#449 · |Δ|max=0.136+ · Common short words, numbers, and everyday technical terms that appear #273 · |Δ|max=0.135+ · Helpful assistant language: giving advice, cautions, step-by-step inst#752 · |Δ|max=0.135− · Mentions of filenames, file paths, and file-related tokens in programm#774 · |Δ|max=0.135+ · Neutral, informative assistant explanations and generic content words #224 · |Δ|max=0.135+ · Text snippets containing numbers, measurements, units, and quantitativ#603 · |Δ|max=0.135+ · Programming source files and code snippets across many languages, espe#654 · |Δ|max=0.135+ · Phrases expressing knowing, realization, certainty, or reported/rumore#122 · |Δ|max=0.134+ · Common tokens and snippets from programming tutorials and code example#642 · |Δ|max=0.134+ · Generic explanatory and encyclopedic prose — common function words and#180 · |Δ|max=0.133+ · Very common short function words and basic punctuation used to connect#751 · |Δ|max=0.133+ · Mathematics and scientific/engineering texts about matrices, linear al#727 · |Δ|max=0.133+ · Common small function words and hedging/temporal phrases used in gener#342 · |Δ|max=0.133+ · Tokens and short word fragments commonly found in encyclopedic, formal#487 · |Δ|max=0.133+ · Formal scientific and medical article language and content (clinical s#132 · |Δ|max=0.132− · Phrases and wording used to introduce examples, links, step‑by‑step in#544 · |Δ|max=0.132− · Graphic sexual content and narratives—often explicit, taboo, or exploi#154 · |Δ|max=0.132+ · Very common short words, small function words, and everyday action/tra#160 · |Δ|max=0.132+ · Scientific and technical text reporting numerical quantities, ranges, #749 · |Δ|max=0.132+ · Very common words and names used in ordinary prose (everyday function # 61 · |Δ|max=0.132+ · Academic scientific research writing in physics and chemistry papers (#445 · |Δ|max=0.132+ · Phrasing about trends, progress, adoption, and changing popularity of # 73 · |Δ|max=0.132+ · Formal academic mathematical writing — theorem/definition statements, #301 · |Δ|max=0.132− · Polite social phrasing: thanks, greetings, and courteous sign-offs use#156 · |Δ|max=0.131+ · Developer-focused API/SDK and cloud infrastructure code snippets, conf#740 · |Δ|max=0.131+ · Very common short function words and small grammatical tokens that app#622 · |Δ|max=0.131+ · Discussion of concurrency and parallelism patterns: threads, locks, po#678 · |Δ|max=0.131+ · Short common words, punctuation, and name tokens found in encyclopedic#260 · |Δ|max=0.131+ · Explanations and step‑by‑step how‑to guidance for programming tasks an#554 · |Δ|max=0.131− · Very common short English function words and handful of generic 'one/s#560 · |Δ|max=0.131− · Practical programming help and how-to code examples, especially Python#442 · |Δ|max=0.131+ · Practical programming how-to content: explanatory steps, examples, cod#636 · |Δ|max=0.131+ · Very high-frequency small words, punctuation, and common function toke#665 · |Δ|max=0.131+ · Words and phrases about places, locations, and asking/indicating where#176 · |Δ|max=0.130− · Short common tokens and punctuation found in technical/math/code and i#517 · |Δ|max=0.130+ · Very common short function words and grammatical connectors that appea#733 · |Δ|max=0.130− · Common discourse connectives and short function words used to link, tr#341 · |Δ|max=0.130+ · Words and short phrases that signal limitation, qualification, or smal#527 · |Δ|max=0.130+ · Code snippets showing conditional checks and guard clauses (if/elif/el# 38 · |Δ|max=0.130+ · Explicit sexual content, often violent or non-consensual, focusing on #773 · |Δ|max=0.130+ · Very common everyday English words and short phrases used in general p#124 · |Δ|max=0.130+ · Tokens that are parts of personal and place names or other capitalized#210 · |Δ|max=0.130+ · Formatting and mention of academic citations, references, and bibliogr#313 · |Δ|max=0.130+ · Text about places, countries, regions, and geographic context (travel,#523 · |Δ|max=0.130+ · Snippets of programming source code showing function/class/method decl#580 · |Δ|max=0.130+ · Very high-frequency common English function words and small grammatica#299 · |Δ|max=0.129+ · Practical programming help: tools, languages, libraries, and how-to gu#300 · |Δ|max=0.129+ · Technical documentation and troubleshooting for software projects: fil#381 · |Δ|max=0.129+ · Programming source code and dependency/import lines from various langu#513 · |Δ|max=0.129− · Boolean/state fields and short structural tokens common in code, confi#407 · |Δ|max=0.129+ · Very common small words, short tokens, and generic phrasing found acro#440 · |Δ|max=0.129+ · Common small words, name fragments, and frequent short tokens found ac# 43 · |Δ|max=0.129+ · Very common short tokens and small function words found across many co#329 · |Δ|max=0.128+ · Technical discussions mentioning operating systems, software versions,#362 · |Δ|max=0.128+ · References to specific dates, years, and time spans in encyclopedic, l#470 · |Δ|max=0.128+ · Mentions of dates, times, and scheduling in code and documentation (ca#728 · |Δ|max=0.128+ · Language discussing nature, naturalness, and what’s innate—human exper#467 · |Δ|max=0.128+ · Code and examples for reading, writing, managing, and inspecting files# 62 · |Δ|max=0.127+ · Short, common words and polite assistant phrasing across Romance langu#150 · |Δ|max=0.127+ · Short code fragments and common programming patterns across multiple l#589 · |Δ|max=0.127+ · Common programming control-flow patterns: condition checks, guard/earl#290 · |Δ|max=0.127+ · Tokens and proper names beginning with the letter sequence K (short wo#771 · |Δ|max=0.127+ · Language and phrasing typical of biomedical and experimental research #444 · |Δ|max=0.126+ · Source-code files and developer-facing documentation (comments, docblo#715 · |Δ|max=0.126+ · Short first‑person and second‑person conversational phrases and common#179 · |Δ|max=0.126+ · Language used in reviews, summaries, and creative-writing guidance—eve#747 · |Δ|max=0.126+ · Phrases and usages that begin with or revolve around the word “under,”#663 · |Δ|max=0.125+ · References to containers, small personal items, and putting/storing ob#679 · |Δ|max=0.125+ · Phrases used in conversational requests, questions, and attentive repl#573 · |Δ|max=0.125+ · Common short function words, punctuation, name fragments, and small fo#211 · |Δ|max=0.125+ · Neutral explanatory assistant responses: concise informational explana#458 · |Δ|max=0.124+ · Polite, hedging, and qualifying language the assistant uses to avoid a#633 · |Δ|max=0.124− · References to official paperwork, signatures, and procedural document #348 · |Δ|max=0.124+ · Phrases that hedge or qualify statements by pointing to variability, i#794 · |Δ|max=0.123− · Tokens found in source code files, especially generated API/SDK code a#400 · |Δ|max=0.123− · Items and short fields from lists, catalogs, indexes, and technical en#438 · |Δ|max=0.123− · Sexualized descriptions of people and erotic roleplay, often explicit #384 · |Δ|max=0.123+ · Common connective and helper phrases used to explain, advise, or descr#162 · |Δ|max=0.123+ · Formal expository and technical writing (encyclopedic, legal, academic#410 · |Δ|max=0.123+ · Common connective and referential words and phrases used in encycloped#204 · |Δ|max=0.122+ · Words and short phrases that link, compare, or indicate relationships #539 · |Δ|max=0.122+ · Marketing and explanatory content about enterprise software, technolog#784 · |Δ|max=0.122− · Text that talks about references, citing, or referring to sources/term#524 · |Δ|max=0.122+ · Legal and constitutional rights and doctrines discussed in court opini# 96 · |Δ|max=0.122+ · Scientific, clinical, and technical reporting language that expresses # 86 · |Δ|max=0.122+ · Explicit sexual roleplay and pornographic dialogue, often involving co#762 · |Δ|max=0.122+ · Python web and backend development code snippets — routing, HTTP handl#785 · |Δ|max=0.122+ · Clear, simple explanatory text—definitions and factual explanations ac#609 · |Δ|max=0.122+ · Technical and scientific descriptions of spatial orientation, angles, #714 · |Δ|max=0.121+ · Phrases that introduce or highlight the central, main, or key part, pr#296 · |Δ|max=0.121+ · Advice, tools, and warnings for keeping people, data, and systems safe#403 · |Δ|max=0.121+ · Technical help and how-to responses in Romance languages, especially p#433 · |Δ|max=0.121+ · User questions and brief conversational prompts asking for help, opini#639 · |Δ|max=0.121+ · Documentation and discussion around software versioning, release notes#681 · |Δ|max=0.121+ · Language about detecting, identifying, and distinguishing things or pa#716 · |Δ|max=0.121+ · Combinatorial probability problems about drawing letters (probabilitie#190 · |Δ|max=0.120+ · Common small function words and connective phrasing used in formal, ex#151 · |Δ|max=0.120+ · Very common short English words and connectors used across many topics#322 · |Δ|max=0.120+ · Explanations and how-to questions: instructional prompts, technical ho#127 · |Δ|max=0.120+ · Common function words and generic connective tokens used in explanator#141 · |Δ|max=0.120+ · Formal explanatory writing in technical, scientific, or instructional #349 · |Δ|max=0.119+ · Practical step‑by‑step programming and developer how‑to instructions (#512 · |Δ|max=0.119+ · Short technical and medical text snippets mentioning specific entities#396 · |Δ|max=0.119+ · Debugging and logging-related code patterns and console output in prog#711 · |Δ|max=0.119+ · Common short function words and connective tokens used across everyday#309 · |Δ|max=0.119+ · Instructions and discussion about making changes, adjustments, or swit# 58 · |Δ|max=0.118+ · Neutral connective and transition words and phrases commonly used in f#502 · |Δ|max=0.118+ · Very common short function words and filler phrases used in everyday, # 93 · |Δ|max=0.118+ · Short words, prefixes, and name fragments that commonly appear in vari# 89 · |Δ|max=0.118+ · Common short function words and everyday connective phrases used to li# 46 · |Δ|max=0.118+ · Practical guidance and code examples for building, training, evaluatin# 85 · |Δ|max=0.118+ · Scientific and academic astronomy/astrophysics writing and technical d#188 · |Δ|max=0.118+ · Formal technical and academic prose introducing or linking ideas, meth#311 · |Δ|max=0.117− · Programming and technical code snippets with numeric identifiers, punc#202 · |Δ|max=0.116+ · Common short discourse words and instructional connectors used in expl#769 · |Δ|max=0.116+ · Mentions of searching and search UI—advice to look things up, search b#809 · |Δ|max=0.116+ · Everyday conversational words and short phrases used in casual narrati#563 · |Δ|max=0.116+ · Phrases and wording that single out, specify, or qualify particular de#336 · |Δ|max=0.116+ · Formal academic and technical writing about improvements, reductions, #443 · |Δ|max=0.116− · C and C++ source/header file structure and boilerplate tokens (namespa#548 · |Δ|max=0.116+ · Numeric tokens and formatted numbers as they appear in tables, scienti#567 · |Δ|max=0.116+ · Source code tokens and boilerplate from programming projects (C/C++ he#750 · |Δ|max=0.116+ · Code patterns handling function exits and early termination (returns, #361 · |Δ|max=0.115+ · Common running-text words and named entities found in encyclopedic, ne#607 · |Δ|max=0.115+ · Texts about firearms, shooting, explosions, and related violent or fir#171 · |Δ|max=0.115+ · Common short substrings, name fragments, and small function words that# 55 · |Δ|max=0.115+ · Common short tokens and assistant-style replies across multiple langua#453 · |Δ|max=0.115+ · Common small function words and generic connective phrasing used in ev#238 · |Δ|max=0.114+ · Tokens and phrases commonly found across diverse English texts—generic#608 · |Δ|max=0.114+ · Short, common words, tokens, and formatting elements that appear acros#388 · |Δ|max=0.114+ · Technical discussions and code about cryptography, authentication, has#402 · |Δ|max=0.114+ · Building and using HTTP APIs: requests, headers, methods, URLs, respon#105 · |Δ|max=0.113+ · German-language text and common German function words, phrases, and as#726 · |Δ|max=0.113+ · Textual mentions of physical sizes, dimensions, and numeric measuremen#398 · |Δ|max=0.113+ · Numeric and statistical reporting in scientific and clinical research # 41 · |Δ|max=0.113+ · In-text citation and numeric reference tokens common in academic, tech#432 · |Δ|max=0.113+ · Common tokens and phrasing found in programming documentation and code#416 · |Δ|max=0.113− · Common short function words and everyday discourse connectors used to #557 · |Δ|max=0.113+ · Biographical and character descriptions in narrative or encyclopedic p#371 · |Δ|max=0.112+ · Phrases and words expressing similarity, sameness, equivalence, or com#172 · |Δ|max=0.112+ · General everyday narrative language and small common words used in inf# 80 · |Δ|max=0.112+ · Language and phrasing typical of judicial opinions and appellate court#253 · |Δ|max=0.112− · Language that describes people's feelings, emotional reactions, and se#627 · |Δ|max=0.112+ · Tokens and phrases commonly found in Wikipedia article category lines #586 · |Δ|max=0.112+ · Common short function words and phrasing used to indicate timing, spee#745 · |Δ|max=0.112− · Questions and answers about finding the n-th term or simplifying polyn#187 · |Δ|max=0.112+ · Common high-frequency words and general-topic nouns used in encycloped#395 · |Δ|max=0.112+ · Formal academic and technical writing about study aims, methods, valid#503 · |Δ|max=0.112+ · Polite, helpful assistant phrasing inviting clarification, follow‑up, #725 · |Δ|max=0.112+ · Neutral explanatory and how‑to prose: generic words and phrases used i#621 · |Δ|max=0.111− · Uses of anonymized NAME placeholders for characters, roles, and entiti#461 · |Δ|max=0.111− · Polite assistant responses about its abilities, limits, and offering h#533 · |Δ|max=0.111+ · Academic biomedical research writing—methods, results, and topic-focus#163 · |Δ|max=0.111+ · Words and names starting with the letters T–Tr (short common tokens, p#731 · |Δ|max=0.111+ · Statements comparing options that highlight benefits, pros, trade‑offs#640 · |Δ|max=0.111+ · Very common English function words and generic discourse connectors us#620 · |Δ|max=0.111+ · Common short words, name fragments, and frequent token patterns that a#656 · |Δ|max=0.110+ · Common short words, name fragments, and place/subject headings that ap#147 · |Δ|max=0.110+ · Text about delays, omissions, missing content, and requests to stop or#232 · |Δ|max=0.110− · Short tokens, symbols, and code-like fragments common in technical, sc#441 · |Δ|max=0.110− · Very common short function words and basic connectors (high-frequency #570 · |Δ|max=0.110+ · Tokens and words beginning with the letters sequence 'ch' (names, scie#628 · |Δ|max=0.110+ · Words and terms that begin with short common letter sequences or prefi#295 · |Δ|max=0.110+ · Very common short words, fragments, and proper-name tokens that serve #739 · |Δ|max=0.110+ · Technical and medical documents referencing standards, codes, and 'I'-#152 · |Δ|max=0.110+ · Very common short function words and small grammatical tokens used acr#754 · |Δ|max=0.110+ · Very common short function words and generic filler tokens used across#791 · |Δ|max=0.110− · Generic short common words and promotional/descriptive phrases found i#550 · |Δ|max=0.109+ · Common words, names, and short tokens that appear in varied encycloped#216 · |Δ|max=0.109+ · Short, everyday words and phrases used to ask questions, make transiti#203 · |Δ|max=0.109+ · Programming examples and discussions about looping and indexing over a#352 · |Δ|max=0.109− · Small numeric and punctuation tokens as they appear in technical, code#383 · |Δ|max=0.109+ · Source code files and programming project snippets across multiple lan#530 · |Δ|max=0.109+ · Code-related testing and test suites: unit/integration tests, test fra#406 · |Δ|max=0.109+ · Very common small words and short connective phrases used in general, # 44 · |Δ|max=0.109+ · Very high‑frequency common words and short grammatical tokens used acr#145 · |Δ|max=0.109+ · Common short function words, hedges, and simple transitional phrases u#359 · |Δ|max=0.109+ · Everyday informal blog and forum writing using common small words, per#511 · |Δ|max=0.108− · Generic function words and connective language used to link ideas, ind#331 · |Δ|max=0.108+ · Step‑by‑step numeric and algebraic problem solving — basic arithmetic,#439 · |Δ|max=0.108+ · Texts about material damage, wear, and protective measures—how things #364 · |Δ|max=0.108− · Text about problems, challenges, and their resolution—describing issue#387 · |Δ|max=0.108+ · General-purpose tokens and headings found in technical, encyclopedic, #676 · |Δ|max=0.108+ · Marketing and local business promotional copy highlighting services, e#339 · |Δ|max=0.108− · Contact and communication information in emails and websites (phone nu#540 · |Δ|max=0.108+ · Source code tokens and common programming syntax across multiple langu#549 · |Δ|max=0.107+ · Text discussing authority, control, governance, and concentration of p# 33 · |Δ|max=0.107+ · Very common English function words, punctuation, and generic boilerpla#227 · |Δ|max=0.107− · Very common small words, connectors, and routine phrases used in infor#389 · |Δ|max=0.107+ · Formal expository writing about projects, products, and processes—effi#351 · |Δ|max=0.107+ · Explanations and guidance about common programming concepts, patterns,#781 · |Δ|max=0.107− · Everyday technical help and troubleshooting language—practical program# 63 · |Δ|max=0.107+ · Common small function words and generic connective phrases used in for#246 · |Δ|max=0.106+ · Academic, scientific, or technical writing and terminology (formal res#318 · |Δ|max=0.106+ · Short tokens and common prefixes (small function words and frequent le#602 · |Δ|max=0.106+ · Concepts and mentions of 'auto' / self-directed or automated systems a#659 · |Δ|max=0.106+ · Language about forming, assembling, or creating things—processes and a#509 · |Δ|max=0.106+ · Words and phrases that describe permanence, lasting change, or habitua#183 · |Δ|max=0.106+ · Tokens and small snippets from programming source code (identifiers, t#363 · |Δ|max=0.106+ · Formal or technical writing in academic, scientific, medical, and tech#591 · |Δ|max=0.106+ · Code examples and snippets from programming Q&A, especially printing/l# 23 · |Δ|max=0.106+ · HTML/CSS front-end layout and styling code (common selectors, properti#155 · |Δ|max=0.106+ · Words, names, and technical terms beginning with the letter F across d#321 · |Δ|max=0.106+ · Language about future events, expectations, and conditional actions (p#667 · |Δ|max=0.106+ · Scholarly, medical, or technical writing and headings—especially acade#266 · |Δ|max=0.106+ · Generic words and phrases common in technical, instructional, or forma#742 · |Δ|max=0.106+ · Phrases that qualify, limit, or reference scope and knowledge (e.g., h#357 · |Δ|max=0.105+ · Front-end web pages and templates: HTML documents with CSS, meta tags,#456 · |Δ|max=0.105− · Formal biomedical and life‑sciences research writing (academic papers:#219 · |Δ|max=0.105+ · Short, common tokens and punctuation that appear in code, technical wr#247 · |Δ|max=0.105+ · Words and phrases expressing size, scale, magnitude, or degree (from v#579 · |Δ|max=0.104+ · Generic English connective and attention-related words and phrases use#468 · |Δ|max=0.104+ · Technical programming help and code-related documentation (developer Q#457 · |Δ|max=0.104+ · Developer troubleshooting and debugging questions about code, errors, #793 · |Δ|max=0.104+ · Very common short words and place/metadata tokens that appear in gener#459 · |Δ|max=0.104+ · Numeric values and formatted numbers as they appear in scientific/medi#720 · |Δ|max=0.104+ · Words and phrases about hiding, covering, or filling spaces—concealmen#136 · |Δ|max=0.104+ · References to dates, years, and time periods in articles, biographies,#719 · |Δ|max=0.104+ · Requests for concise summaries, overviews, or brief explanatory how‑to# 99 · |Δ|max=0.104+ · Formal expository and academic prose—general-purpose words and phrases#618 · |Δ|max=0.104+ · Very common function words, short connectors, and routine punctuation # 21 · |Δ|max=0.103+ · Everyday connective language and hedging used in narratives, explanati# 48 · |Δ|max=0.103+ · Common tokens and formatting found in diverse written documents—short #254 · |Δ|max=0.103+ · Very common small function words and short grammatical connectors that#486 · |Δ|max=0.103+ · Language used in legal, business, and real‑estate agreements and negot#564 · |Δ|max=0.103+ · Common short function words and small phrases about movement/location #600 · |Δ|max=0.103+ · Code patterns for attaching and handling event callbacks/listeners in #664 · |Δ|max=0.103+ · Technical and scientific writing about resources, inputs, sources, tar# 74 · |Δ|max=0.103+ · Questions and code about relational databases and SQL (queries, schema#743 · |Δ|max=0.103− · Short tasks asking the model to judge everyday behaviors as “it's okay#408 · |Δ|max=0.103+ · Phrases and small words used to instruct, urge, or describe actions an#777 · |Δ|max=0.103+ · Statements emphasizing prioritizing people’s safety, well‑being, and a#694 · |Δ|max=0.102+ · Language commonly used in organizational or technical support, promoti#316 · |Δ|max=0.102+ · Scholarly academic writing about philosophy, social sciences, and publ#724 · |Δ|max=0.102+ · Words and phrases expressing continuation, ongoingness, or that someth#450 · |Δ|max=0.102+ · Phrases and wording writers use to comment on how things seem, look, o#447 · |Δ|max=0.102+ · Neutral explanatory language used in advice, reports, or policies desc#626 · |Δ|max=0.102− · Promotional and informational web copy: marketing, calls-to-action, an# 84 · |Δ|max=0.102+ · Very common small words and grammatical tokens used in neutral explana# 13 · |Δ|max=0.101− · Very common short words and grammatical tokens that appear in ordinary#226 · |Δ|max=0.101+ · Generic high-frequency English words and small grammatical tokens used#426 · |Δ|max=0.101+ · Everyday conversational words and short phrases used to make, contrast#233 · |Δ|max=0.101+ · Everyday informational and connective words and phrases used in articl#109 · |Δ|max=0.101+ · Generic language patterns and common function words used in everyday e#314 · |Δ|max=0.101− · Formal academic and technical prose (scientific papers, patents, and t#764 · |Δ|max=0.101+ · Writing about celebrities, fame, and cultural impact—tabloid-style cov#420 · |Δ|max=0.100+ · Academic biomedical research text discussing molecular pathways, exper# 76 · |Δ|max=0.100+ · Mentions of dates and times in factual or informational writing (calen#272 · |Δ|max=0.100+ · Proper names, company/place names, and common short function words fou#525 · |Δ|max=0.100+ · Academic and clinical biomedical writing about diseases, pathology, tr#125 · |Δ|max=0.100+ · News and legal writing about criminal charges, arrests, prosecutions, #537 · |Δ|max=0.100+ · Text about working, collaboration, and whether things or people functi#546 · |Δ|max=0.099− · Biomedical texts about cancer: tumors, pathology, metastasis, biomarke#605 · |Δ|max=0.099+ · Text about newness—introducing, creating, or adapting to something new# 81 · |Δ|max=0.099+ · Text about clothing, fashion items, and sewing/craft projects related #490 · |Δ|max=0.099+ · Common small connective and function words found in technical, legal, #430 · |Δ|max=0.099− · Greetings, salutations, and brief polite openings used to start conver# 56 · |Δ|max=0.099+ · Phrases and reasoning courts use to describe evidentiary standards, bu#195 · |Δ|max=0.099+ · Scholarly biomedical and health-science writing about physiological me#693 · |Δ|max=0.099+ · Text about organizations, groups, and their structures, history, and a#646 · |Δ|max=0.098+ · Very common short function words and generic connective tokens used ac#611 · |Δ|max=0.098+ · Words and phrases about places, distances, and how near or far things #585 · |Δ|max=0.098− · Uses of the root “act/activ/active” and related action-language across#257 · |Δ|max=0.098+ · Common small connective words, punctuation, and short function tokens #674 · |Δ|max=0.098+ · Words and text snippets featuring 'comp'/'com' beginnings—terms like c#345 · |Δ|max=0.097+ · Common small function words and short connective phrases used to link,#577 · |Δ|max=0.097+ · Formal scientific and technical writing patterns for reporting measure# 94 · |Δ|max=0.097− · Tropes and elements common in genre fiction summaries and plot descrip#401 · |Δ|max=0.096− · Very common short words and tokens that start or appear in assistant r#691 · |Δ|max=0.096+ · Short pieces of technical text focused on numbers: arithmetic/math pro#138 · |Δ|max=0.095+ · Very common small English words and generic connective/qualification p# 54 · |Δ|max=0.095+ · Practical command-line, scripting, packaging, and system administratio#198 · |Δ|max=0.095− · Common short grammatical words, punctuation, and formatting tokens tha#281 · |Δ|max=0.095+ · Neutral, informative assistant text: explanatory how‑tos, factual answ#284 · |Δ|max=0.095+ · Web front-end coding: HTML/JavaScript (DOM manipulation, jQuery, UI wi#466 · |Δ|max=0.095+ · Very common short English words and little connecting phrases used acr# 51 · |Δ|max=0.095+ · Very common short words and generic connective phrasing found across m#495 · |Δ|max=0.095+ · Talk about time—how people spend, have, or take short periods (sleep, #477 · |Δ|max=0.095− · Short, common English words and function tokens that glue sentences to#552 · |Δ|max=0.095− · Content about teaching, learning, and training—courses, tips, programs#718 · |Δ|max=0.095+ · Language about achieving goals, success, effort, and ways to accomplis#270 · |Δ|max=0.095+ · Common little words and connective phrases used in neutral explanatory#209 · |Δ|max=0.094− · Polite assistant language offering help, follow-ups, and brief self-de#394 · |Δ|max=0.094+ · Standard ethical, funding, and boilerplate statements found in academi# 95 · |Δ|max=0.094+ · Technical networking and telecommunications topics: protocols, configu#702 · |Δ|max=0.094+ · Words, names, and tokens containing the letters Z/zo/zu/zen (proper na#576 · |Δ|max=0.094− · User prompts and assistant responses for producing written content or #671 · |Δ|max=0.094+ · Boilerplate open-source license header text found at the top of source#643 · |Δ|max=0.093+ · Common short function words and everyday connective language used in o#333 · |Δ|max=0.093+ · Common uses of the short function word “at” and related small temporal#174 · |Δ|max=0.093+ · Everyday expressions of personal reaction, interest, and polite announ#489 · |Δ|max=0.093+ · Very common small function and connective words used to link ideas, in#746 · |Δ|max=0.093+ · Casual to aggressive swear words, insults, and explicit bodily-excreti#522 · |Δ|max=0.093+ · Common small function words and short connective tokens (especially us#757 · |Δ|max=0.093+ · Auto-generated source file headers, comments, and the surrounding code#256 · |Δ|max=0.092− · Formal expository and instructional writing using common function word#161 · |Δ|max=0.092+ · Programming texts about handling, searching, validating, and manipulat# 88 · |Δ|max=0.092+ · Text about military and naval history, units, ships, deployments, and #335 · |Δ|max=0.092− · Generic everyday language and common instructional/commercial phrasing#504 · |Δ|max=0.092+ · Common short function words and casual conversational phrases centered#123 · |Δ|max=0.092+ · Place names and geographic references in encyclopedia-style and news t#144 · |Δ|max=0.092+ · HTML and web frontend markup commonly found in documentation, navigati# 30 · |Δ|max=0.092− · Explanatory, instructional, and informational phrasing used in how‑tos# 82 · |Δ|max=0.091+ · Runtime and compile-time error messages, logs, and stack traces from p#234 · |Δ|max=0.091+ · Very common short function words and simple transition/time phrases fo#191 · |Δ|max=0.091+ · Inline citation markers and reference-bracket tokens commonly used in #236 · |Δ|max=0.091− · Formatting and numbering tokens common in academic, technical, or code#565 · |Δ|max=0.090+ · Text about rating, scoring, ranking, and reviews (numerical or categor#498 · |Δ|max=0.090+ · Common small function words and tokens used in technical, numerical, a#177 · |Δ|max=0.090− · Medical and pharmaceutical discussions about drugs: their uses, dosing#369 · |Δ|max=0.090+ · Technical and programming text: code snippets, math/formal definitions#365 · |Δ|max=0.089+ · Encyclopedic historical writing—overviews and discussions of past peri#327 · |Δ|max=0.089+ · Section headings and standard scholarly article text (methods, results#356 · |Δ|max=0.089+ · Common small grammatical words, connective phrases, and boilerplate wo#285 · |Δ|max=0.089+ · Scientific and technical writing about chemicals, materials, and labor#707 · |Δ|max=0.089+ · Phrases and small function words used in everyday explanations, instru#244 · |Δ|max=0.089− · Common small words and general-purpose phrasing used in everyday expla#252 · |Δ|max=0.089+ · Very common short function words and formatting tokens that appear acr#289 · |Δ|max=0.089+ · Technical and engineering descriptions of fluid, thermal, and mechanic#485 · |Δ|max=0.089+ · Comparative and contrast language used in technical, scientific, and f# 59 · |Δ|max=0.088+ · Technical descriptions of industrial organic chemistry: synthesis, pro#480 · |Δ|max=0.088+ · Assistant statements noting its limits, lack of consciousness, or inab#614 · |Δ|max=0.088+ · Common small grammatical words and short instruction-related phrasing #192 · |Δ|max=0.088− · Clinical and biomedical texts reporting blood-based laboratory measure#606 · |Δ|max=0.088+ · Words and phrases related to passing, transmitting, sharing, or handin# 49 · |Δ|max=0.088+ · News and legal reporting about crimes, arrests, violent incidents, dea#698 · |Δ|max=0.088+ · Phrases and wording that describe or evaluate ease, simplicity, readin# 66 · |Δ|max=0.087+ · Casual everyday language and personal-expression tokens used in inform#328 · |Δ|max=0.087+ · Discussions of gender, sexual orientation, and experiences of marginal#479 · |Δ|max=0.087+ · Short words, name fragments, and common substrings that appear across # 7 · |Δ|max=0.087+ · Formal technical and patent-style descriptions of mechanical/electrica#324 · |Δ|max=0.087− · Generating poems, song lyrics, and other creative lyrical writing (inc# 57 · |Δ|max=0.087+ · Clinical and medical research writing: case reports, trials, diagnoses#360 · |Δ|max=0.087+ · Discussions and code about colors, color properties, and background st#596 · |Δ|max=0.087+ · Technical and scientific writing (academic papers, methods, and discus#376 · |Δ|max=0.087+ · Common words and phrasing found in formal, technical, and academic doc#111 · |Δ|max=0.087+ · Words and phrases that mark time, durations, and temporal references i#240 · |Δ|max=0.087+ · Problems and notations involving numbers and numeric sequences in tech# 69 · |Δ|max=0.086+ · Supportive self-help and spiritual guidance encouraging personal growt#393 · |Δ|max=0.086+ · Words and phrases about leaving, staying, and keeping or storing thing#741 · |Δ|max=0.086+ · Encyclopedic and local content about German-speaking countries — place#672 · |Δ|max=0.086+ · Phrases and language used to ask for technical help, troubleshooting, #101 · |Δ|max=0.086+ · Content about online promotion, social media, websites, and digital ma#562 · |Δ|max=0.086+ · Practical workplace and project management guidance—planning, processe#775 · |Δ|max=0.086+ · Instructions and questions asking to round or format numbers to a spec#675 · |Δ|max=0.086+ · Text about dividing, splitting, or grouping things into parts, categor#753 · |Δ|max=0.086+ · Markers, names, and links used in online discussion replies and quoted#312 · |Δ|max=0.086− · Structured technical text and common degenerate/repetitive token patte#149 · |Δ|max=0.086+ · Very common short function words, small grammatical tokens, and everyd#688 · |Δ|max=0.086+ · Words and phrases related to play, fun, joking, and playful behavior i#662 · |Δ|max=0.085+ · Language and boilerplate commonly found in judicial opinions and appel#302 · |Δ|max=0.085+ · Biographical and news text describing people's names, roles, leadershi#689 · |Δ|max=0.085+ · Text about reading, recommending, referencing, or linking to books/art#644 · |Δ|max=0.085+ · Uses of the word “from” and other common small function words in headi#682 · |Δ|max=0.085+ · Text about immigration, visas, refugees, and legal/policy issues aroun# 60 · |Δ|max=0.085+ · Common function words and generic informational phrasing found in ency# 19 · |Δ|max=0.085+ · News and reference text about elections, political officeholders, camp#471 · |Δ|max=0.085+ · Text about light, brightness, and optical/vision topics (sunlight, ill#528 · |Δ|max=0.085+ · General English text and encyclopedic content containing many common s#354 · |Δ|max=0.084+ · General conversational language and narrative tokens used in varied we#317 · |Δ|max=0.084+ · Formal legal opinion language describing court proceedings, parties, a#551 · |Δ|max=0.084+ · Very common small words and generic content words used in everyday pro#723 · |Δ|max=0.084+ · Words and phrases that start with or contain the “col-/coll-” root, es#704 · |Δ|max=0.084+ · Algebra and calculus problems focused on simplifying power/exponent ex# 87 · |Δ|max=0.084+ · Text about buying, selling, shipping, and practical commercial transac#706 · |Δ|max=0.084+ · Text discussing properties, characteristics, or named attributes of th#786 · |Δ|max=0.084− · Flutter and Dart UI code patterns and typical widget/state usage examp#571 · |Δ|max=0.083+ · Math word‑problem tokens about integers: computing GCD/LCM, digits, re# 31 · |Δ|max=0.083+ · Very common small words and generic phrasing used in informational, ac#422 · |Δ|max=0.083+ · Frequent small function words and common connective phrases used acros#507 · |Δ|max=0.083+ · Descriptions and references to methods, approaches, paths, and step-by#261 · |Δ|max=0.083+ · Programming source code tokens and common syntax patterns found in C-l#637 · |Δ|max=0.083+ · Phrases and usages centered on 'case' and situational/contextual frami#569 · |Δ|max=0.083+ · Texts and phrases about homes, living situations, and the concept of l#197 · |Δ|max=0.083+ · Common small function words and generic tokens found in normal prose, #201 · |Δ|max=0.082+ · Text about romantic and sexual attraction, relationships, and social/s#230 · |Δ|max=0.082+ · Code examples and how-to guidance for data analysis plotting and dataf#481 · |Δ|max=0.082+ · Short conversational markers and tasks for confirming factual consiste#710 · |Δ|max=0.082+ · Very common short function words and punctuation used as neutral conne#130 · |Δ|max=0.082+ · Methods and equipment details for laboratory experiments and biomedica#241 · |Δ|max=0.082+ · Short technical and informational phrases common in documentation, cod# 72 · |Δ|max=0.082+ · Casual multilingual social-media style text (emoji-rich, noisy or non-#280 · |Δ|max=0.082− · Tokens and phrasing related to personal names, name placeholders, and #185 · |Δ|max=0.081+ · Text from web pages and encyclopedic articles focused on South Asian t#134 · |Δ|max=0.081+ · Text and instructions for graphical user interfaces—buttons, menus, di#310 · |Δ|max=0.081+ · Phrases and short words that express ability, inability, possibility, #668 · |Δ|max=0.081+ · Topics about cleaning, washing, hygiene, and household personal-care p#237 · |Δ|max=0.081+ · Encyclopedic biographical text about people's lives, careers, dates, a#451 · |Δ|max=0.081+ · Positive, reassuring, or evaluative phrasing used in everyday advice, # 12 · |Δ|max=0.081+ · Very common English function words and generic formal prose tokens tha#434 · |Δ|max=0.081− · Informational and technical documentation describing what tools/produc#473 · |Δ|max=0.081+ · Questions and short phrases asking for reasons or explanations (everyd#146 · |Δ|max=0.080+ · Phrases and expressions that refer to time, duration, or temporal poin# 37 · |Δ|max=0.080+ · Common, high-frequency words and connective phrasing found in formal/t#235 · |Δ|max=0.080+ · Very common short function words and generic boilerplate phrasing foun#206 · |Δ|max=0.080+ · Words and phrases that mark order, timing, or novelty (e.g., 'first', # 34 · |Δ|max=0.080+ · Tokens that commonly appear in news or articles when mentioning people#759 · |Δ|max=0.080+ · Mentions of a particular area, field, location, or 'on the ground' con#555 · |Δ|max=0.080+ · Tokens and phrases common in algebraic arithmetic problems — numeric a#297 · |Δ|max=0.080+ · Markup and web/service configuration files and tags (XML/HTML/XSLT/XPa#677 · |Δ|max=0.080− · Polite well-wishes and closing phrases used by writers or assistants (#344 · |Δ|max=0.079+ · Text about visual art, artists, exhibitions, museum collections, and i#148 · |Δ|max=0.079+ · Common small function words and connective phrases used in questions, #265 · |Δ|max=0.079+ · Common Russian-language assistant and informational prose: neutral con#332 · |Δ|max=0.079+ · Development questions and code snippets about .NET/C# application UI, #469 · |Δ|max=0.079− · Algebra word-problem language: defining variables/functions (often sta#193 · |Δ|max=0.079+ · Algebraic arithmetic and simplification problems—combining exponents, #455 · |Δ|max=0.079+ · Short arithmetic and algebra problems and requests (basic integer/frac#661 · |Δ|max=0.079+ · Expressions of affection, love, and enjoying shared leisure or compani#337 · |Δ|max=0.078+ · Mentions of people’s ages, age groups, and age-related demographic inf#409 · |Δ|max=0.078− · Short, common words and boilerplate phrases that appear across many co# 26 · |Δ|max=0.078+ · Formal academic and scientific writing—methods, procedures, and common#638 · |Δ|max=0.078+ · References to open-source/free software, licensing, and related projec#142 · |Δ|max=0.077+ · Everyday conversational and instructional language—polite offers, requ#578 · |Δ|max=0.077+ · Instructional and how-to text framing (phrases that introduce steps, e#119 · |Δ|max=0.077+ · News-style reporting language and boilerplate (quotations, attribution#604 · |Δ|max=0.077+ · Tokens and punctuation common in code, markup, and technical documenta#553 · |Δ|max=0.076+ · Common short function words and simple temporal/transition phrases use#748 · |Δ|max=0.076+ · Formatting and structural markers used in technical and academic docum#223 · |Δ|max=0.076+ · Formal scientific and medical research writing reporting experimental #229 · |Δ|max=0.076+ · Texts about funding, donations, and programs to support communities, n#276 · |Δ|max=0.076+ · Common high-frequency function words and generic connective phrasing t#508 · |Δ|max=0.076+ · Phrases about remembering, memories, and reminders (nostalgia, 'rememb#106 · |Δ|max=0.076+ · Common short words, pronouns, punctuation, and everyday discourse mark#465 · |Δ|max=0.076+ · Language used in clinical research papers to report patient outcomes, #669 · |Δ|max=0.076+ · Short conversational prompts and responses typical of everyday chat: q#547 · |Δ|max=0.075+ · Descriptions, guidance, and assessments of physical movement, posture,#534 · |Δ|max=0.075+ · Wikipedia-style encyclopedic article text (biographies, places, organi#392 · |Δ|max=0.075− · Concise, practical Stack Overflow–style programming help: code fixes, #543 · |Δ|max=0.075+ · Words and names containing the letter sequence “ol/ol-” (proper names,#696 · |Δ|max=0.075+ · Tokens and phrases used when the assistant begins or is asked to compl#766 · |Δ|max=0.075+ · Technical content about maps, coordinates, and location services (mapp#501 · |Δ|max=0.074+ · Language and phrases used to present, compare, and make choices or dec#267 · |Δ|max=0.074+ · Very common short function words and small connective tokens appearing#574 · |Δ|max=0.074+ · Academic biomedical research and grant-writing language describing stu#418 · |Δ|max=0.074+ · Generic common words, names, and short topical tokens found in diverse#153 · |Δ|max=0.074+ · Common everyday English words and short phrases used to connect, refer#421 · |Δ|max=0.074+ · Phrases and verbs centered on 'take'—actions of taking, undertaking, a#277 · |Δ|max=0.074+ · Phrases and words used to introduce, structure, and state aims or tran#624 · |Δ|max=0.074+ · Requests for written professional content and structured summaries (co#158 · |Δ|max=0.074+ · Technical descriptions of materials, their synthesis, properties, and #415 · |Δ|max=0.074+ · Casual online comments and forum-style conversational fragments (posts# 5 · |Δ|max=0.073+ · Cooking and food-related recipe content (instructions, ingredients, pr#143 · |Δ|max=0.073+ · Academic and public-health research writing about mental/physical heal# 90 · |Δ|max=0.073+ · News-style reporting and analysis of international conflicts, crises, #632 · |Δ|max=0.073− · Phrases and words common in assistant replies and user instruction-fol#655 · |Δ|max=0.073+ · Requests and outputs for basic integer arithmetic tasks—especially fin#275 · |Δ|max=0.073+ · First-person expressions and conversational phrasing used to state bel# 78 · |Δ|max=0.073+ · Common sentence-level connective and transitional words/phrases used t#436 · |Δ|max=0.073+ · Mentions of towns, cities, local places, and related local government #597 · |Δ|max=0.073− · Short, common words and connectors used in simple factual explanations#425 · |Δ|max=0.073+ · Questions and answers about programming problems, debugging, and pract#717 · |Δ|max=0.073+ · Methods and materials descriptions in scientific/biomedical papers (ex#598 · |Δ|max=0.072+ · Words and phrases about being part of, included in, or belonging to a #760 · |Δ|max=0.072+ · Algebraic simplification of numeric expressions with many square roots#695 · |Δ|max=0.072+ · Encyclopedic and administrative content about East Asian places, peopl#599 · |Δ|max=0.072+ · Generic technical and documentation language: short function words, co#117 · |Δ|max=0.072− · Short, commonly used words and phrases that appear in explanatory, gra#670 · |Δ|max=0.072+ · Text and citation formatting from U.S. judicial opinions and appellate#245 · |Δ|max=0.072+ · Technical discussion and documentation about creating, processing, and# 11 · |Δ|max=0.071+ · Formal scientific and technical prose describing methods, materials, m#705 · |Δ|max=0.071− · Statistical methods and reporting of outcomes, group comparisons, and #315 · |Δ|max=0.071+ · Text that reports or references numerical proportions, percentages, an#610 · |Δ|max=0.071+ · Texts about ordering and sorting things—instructions, questions, or co#658 · |Δ|max=0.071+ · Oligonucleotide sequences and PCR/qPCR primer details in molecular bio#165 · |Δ|max=0.071+ · JavaScript and frontend web development code examples and troubleshoot#397 · |Δ|max=0.070+ · Multilingual assistant conversation tokens and polite/helpful phrasing#631 · |Δ|max=0.070− · Short math and arithmetic problem prompts and small-function words com#263 · |Δ|max=0.070+ · Phrases and passages expressing uncertainty, lack of knowledge, or not#355 · |Δ|max=0.070+ · Common short function words and connective phrases used to start sente# 22 · |Δ|max=0.069− · Common everyday words, small grammatical tokens, and routine phrases f# 75 · |Δ|max=0.069+ · Journalistic/news article language and formatting (headlines, bylines,#583 · |Δ|max=0.069+ · Phrases and tokens from clinical and biomedical research reports descr#306 · |Δ|max=0.069+ · Website and forum UI boilerplate and community engagement prompts (reg#343 · |Δ|max=0.069+ · Language and formatting commonly found in judicial opinions and appell#370 · |Δ|max=0.069+ · Very common everyday words, small grammatical tokens, and routine UI/c#366 · |Δ|max=0.069+ · Common small function words and simple discourse connectors used to st#108 · |Δ|max=0.068+ · Descriptions and discussions of weather, temperature, seasons, and pla#435 · |Δ|max=0.068+ · Language for comparing, ranking, or calling something the most/least (#417 · |Δ|max=0.068+ · Technical programming Q&A posts: debugging, how-to questions, code sni# 98 · |Δ|max=0.068+ · Calling out or citing figures, tables, and other labeled visuals in ac#454 · |Δ|max=0.068− · Clinical and biomedical research writing describing patients, study pa#319 · |Δ|max=0.068+ · Phrases and wording typical of academic research articles (introducing#687 · |Δ|max=0.067+ · Assistant safety and refusal boilerplate, plus user jailbreak prompts #228 · |Δ|max=0.067+ · Very common short words, punctuation, and structural tokens that appea#647 · |Δ|max=0.067− · Descriptions, reviews, and sensory/production discussion of alcoholic #114 · |Δ|max=0.067+ · Text about families, parenting, and close kin relationships (parents, #121 · |Δ|max=0.067+ · Text about jobs, hiring, careers, and professional roles (postings, ap#709 · |Δ|max=0.067+ · Practical programming help and troubleshooting for developers (Stack O#493 · |Δ|max=0.067+ · Tokens and punctuation patterns common in technical, academic, and pat#293 · |Δ|max=0.066+ · Common connective and formal phrasing used to state reasons, bases, an#182 · |Δ|max=0.066+ · Technical discussion and code for 2D/3D graphics and game development #178 · |Δ|max=0.066+ · Common short function words and formatting tokens found in technical, #572 · |Δ|max=0.066+ · Programming problems and solutions focused on algorithms, data structu#431 · |Δ|max=0.066+ · Sections of scientific papers describing X‑ray crystallography experim#541 · |Δ|max=0.066+ · Short math problems and algebraic simplifications (school-level arithm#424 · |Δ|max=0.066+ · Common small function words and everyday connective language found acr#634 · |Δ|max=0.066+ · Academic and news reporting of population-level statistics, prevalence# 9 · |Δ|max=0.066+ · Legal and contractual texts about liability, corporate/government enti#595 · |Δ|max=0.065+ · Tokens and phrases common to personal blogs and informal dated posts (#690 · |Δ|max=0.065+ · Text about industrial or chemical production, manufacturing processes,#170 · |Δ|max=0.065+ · Formal expository and instructional prose using common connective word#323 · |Δ|max=0.065+ · Java server-side development: Spring/Servlet-based web apps, framework#126 · |Δ|max=0.065+ · Encyclopedic writing about historical places, people, institutions, an#307 · |Δ|max=0.065− · Prompts and assistant reply openings that instruct the model to only c#287 · |Δ|max=0.065+ · Mentions of money, prices, costs, and financial figures or pricing-rel#377 · |Δ|max=0.065+ · Formal patent and technical-document language describing prior art, pr#239 · |Δ|max=0.065+ · Go programming language source code patterns and idioms (including gen#181 · |Δ|max=0.065− · Instruction prompts for short, structured responses and evaluations (m#258 · |Δ|max=0.065+ · Common small function words and everyday connective phrases found thro#515 · |Δ|max=0.065− · Short math problem language: elementary arithmetic, integer operations#427 · |Δ|max=0.064+ · References to figures, figure panels, image captions, and related labe# 6 · |Δ|max=0.064+ · News and reports about finance, markets, and government fiscal/economi#703 · |Δ|max=0.064+ · Common short/high-frequency tokens and simple math problem wording—eve# 53 · |Δ|max=0.064+ · Descriptions and listings of buildings and homes, including architectu#113 · |Δ|max=0.064+ · Formal legal and regulatory texts with statutory citations, section ma#701 · |Δ|max=0.064+ · Phrases about trying, attempting, or encouraging people to try (effort#140 · |Δ|max=0.064+ · Automotive content: cars, manufacturers, models, specifications, and r#582 · |Δ|max=0.064− · Everyday user questions and requests asking for explanations, recommen#625 · |Δ|max=0.064− · Phrases and tokens common in academic and technical papers—section hea#374 · |Δ|max=0.064− · Typical assistant responses to user questions: factual answers, how‑to# 25 · |Δ|max=0.063+ · Boilerplate open‑source software license and copyright headers found i# 47 · |Δ|max=0.063+ · Mentions of geographic places, location words, and travel/encyclopedic#168 · |Δ|max=0.063+ · Text about public events, meetings, conferences, festivals, and their #135 · |Δ|max=0.063+ · Product descriptions and technical details for consumer mobile and hom#378 · |Δ|max=0.063+ · Words and phrases used to describe geographic locations, directions, a#520 · |Δ|max=0.063− · Academic biomedical and life-science prose reporting experimental meth# 92 · |Δ|max=0.063+ · Auto-generated software documentation and its HTML navigation/structur#545 · |Δ|max=0.063+ · Server‑side web framework code and common patterns for building routes#157 · |Δ|max=0.062+ · Low-level system and device driver source code and documentation (kern#496 · |Δ|max=0.062+ · Text about awards, honors, recognitions, and formal distinctions given# 68 · |Δ|max=0.062+ · Common short function words, connectors, and everyday filler phrasing #107 · |Δ|max=0.062+ · Phrases where the assistant explains its identity, capabilities, limit#338 · |Δ|max=0.061+ · Common small function words, connective phrases, and generic scaffoldi#492 · |Δ|max=0.061+ · Formal informational and organizational prose found in articles, abstr#199 · |Δ|max=0.061+ · Structural elements and references in academic scientific papers (sect# 28 · |Δ|max=0.061+ · Content about movies, TV shows, actors, production credits, and relate#350 · |Δ|max=0.061+ · C++ library and systems C++ source code — template/Boost/Eigen/Chromiu#538 · |Δ|max=0.061− · Phrases expressing apology, regret, or polite refusal (including empat#283 · |Δ|max=0.061+ · Phrases and prompts for asking, offering, or producing written content#692 · |Δ|max=0.060+ · Local and national public-service, government, and institutional infor# 42 · |Δ|max=0.059+ · Text about educational institutions, academic credentials, and school-#242 · |Δ|max=0.059− · User prompts and roleplays trying to override assistant rules—jailbrea#196 · |Δ|max=0.059+ · Laboratory and industrial chemical synthesis procedures and purificati#372 · |Δ|max=0.059+ · Technical discussions and code examples about low-level systems progra#581 · |Δ|max=0.059+ · Assistant messages about language abilities, translations, and switchi#683 · |Δ|max=0.058+ · Headings, captions, and numeric citations found in U.S. court opinions#763 · |Δ|max=0.058+ · Scientific discussion of oxidative stress, reactive oxygen species, an#120 · |Δ|max=0.058+ · Common short function words and connective phrases used across news, a#463 · |Δ|max=0.058− · Very high‑frequency short words and boilerplate language that appear a# 15 · |Δ|max=0.058+ · Copyright and open-source license header boilerplate found at the top #249 · |Δ|max=0.057+ · Formal academic and scientific writing—especially biomedical, biologic# 39 · |Δ|max=0.057+ · Discussion of cellular protein functions and pathways—binding, modific#159 · |Δ|max=0.057+ · Language and structure of appellate and trial court opinions — procedu#334 · |Δ|max=0.057+ · Deciding whether a short summary is factually consistent with a source#542 · |Δ|max=0.056− · Short algebra and arithmetic word problems focusing on simple polynomi#205 · |Δ|max=0.056+ · Mentions and markup for figures, tables, images, and supplementary mat#765 · |Δ|max=0.056− · Tokens from SVG/vector icon path data and other vector-graphics markup#303 · |Δ|max=0.055+ · Text from judicial opinions and legal case citations (court reporter r# 24 · |Δ|max=0.055+ · Medical and surgical descriptions of human anatomy, injuries, vascular#713 · |Δ|max=0.054− · Text strings, metadata, and file formats used for software localizatio#189 · |Δ|max=0.054+ · Health and nutrition content about diet, nutrients, body weight, and r#405 · |Δ|max=0.054+ · Combinatorics word problems asking for probabilities of drawing specif#521 · |Δ|max=0.054+ · Text about peoples, cultures, and their histories—including indigenous# 36 · |Δ|max=0.053+ · Generic website boilerplate and legal/disclaimer copy found on commerc#274 · |Δ|max=0.053+ · Scholarly neuroscience texts about brain systems, sensory/motor functi#532 · |Δ|max=0.053+ · Place names and postal/address fragments found in formal documents (ad#568 · |Δ|max=0.053+ · Short algebra and arithmetic word-problems: solving simple linear equa#271 · |Δ|max=0.053− · Short math problem prompts asking for specific numeric results (remain#373 · |Δ|max=0.053− · Information about chemical hazards and practical safety precautions fo#208 · |Δ|max=0.052+ · Formal academic and scientific paper text (research articles, abstract#214 · |Δ|max=0.052+ · Text about transportation and travel infrastructure (stations, routes,#514 · |Δ|max=0.051+ · Scientific and medical research text about human and animal anatomy, d# 16 · |Δ|max=0.051− · Promotional ‘About the Company’ introductions for chemical manufacture#488 · |Δ|max=0.051+ · Questions that compute probabilities of drawing specific letters, sequ#320 · |Δ|max=0.050+ · Short algebra problems that simplify polynomials, extract numeric coef#133 · |Δ|max=0.050+ · Language and formatting typical of judicial opinions and legal case ci#288 · |Δ|max=0.050+ · Website and blog content elements—sidebars, footers, metadata, listing#340 · |Δ|max=0.050− · Everyday conversational and informational assistant language—common wo# 79 · |Δ|max=0.049+ · Formal academic and biomedical research writing: methods, results, qua# 20 · |Δ|max=0.049+ · Legal court opinions and filings focusing on procedural history, motio#462 · |Δ|max=0.049+ · Informal spoken and quoted language in news, interviews, and media (tr#613 · |Δ|max=0.049+ · Encyclopedic and news-style content about places, people, and institut#412 · |Δ|max=0.049+ · Solving and simplifying algebra problems about polynomials, roots, fac# 70 · |Δ|max=0.049+ · Travel and tourism descriptions: places, attractions, outdoor activiti#129 · |Δ|max=0.049+ · Text about energy and environmental issues, especially renewable energ#304 · |Δ|max=0.048+ · Mobile app user interface and lifecycle development (Android and iOS l# 29 · |Δ|max=0.048+ · Encyclopedic entries and metadata about music (artists, albums, perfor#561 · |Δ|max=0.048− · Questions and problems about divisibility, remainders, and numeric com#115 · |Δ|max=0.048− · Academic and clinical microbiology/virology texts about infectious pat#347 · |Δ|max=0.048+ · Methods and experimental descriptions from biomedical research article#213 · |Δ|max=0.047+ · High-entropy technical identifiers and encoded data found in software,#478 · |Δ|max=0.047+ · Instructions and tasks about rounding or computing numerical results t#736 · |Δ|max=0.047+ · Positive customer reviews praising products, services, and staff# 40 · |Δ|max=0.047+ · Technical descriptions of molecular biology and genomics methods, reag# 52 · |Δ|max=0.047+ · Standard open-source license header text and warranty/liability discla#617 · |Δ|max=0.047+ · Language and citation conventions found in U.S. appellate and judicial#474 · |Δ|max=0.047− · LaTeX and Math typesetting in academic and technical documents (equati#721 · |Δ|max=0.047+ · Questions and code for using Excel and Office automation (VBA, macros,#112 · |Δ|max=0.046+ · Methods, sampling, and reporting language from medical and public-heal#735 · |Δ|max=0.045+ · Code and discussion about building graphical user interfaces — widgets#494 · |Δ|max=0.045+ · Questions and calculations converting or computing quantities, especia# 32 · |Δ|max=0.045+ · Reporting and describing statistical methods, results, and significanc#382 · |Δ|max=0.045+ · Formal, encyclopedic and literary prose from older printed texts—biogr#768 · |Δ|max=0.044+ · Website and document placeholder text and template filler (e.g., Lorem#708 · |Δ|max=0.044+ · Text about fights, combat, and competitive conflict—especially combat #164 · |Δ|max=0.043+ · Scientific discussions of immune system components, immune responses, #761 · |Δ|max=0.043+ · Formal legal documents and appellate court opinions (case captions, ci#212 · |Δ|max=0.043+ · Structural words and phrases found in Wikipedia articles and disambigu#220 · |Δ|max=0.043+ · Questions and exercises that compare, rank, or sort numbers (find smal#516 · |Δ|max=0.043+ · Medical and biological topics related to pregnancy, childbirth, newbor#368 · |Δ|max=0.042+ · Short math question tokens and wording found in elementary arithmetic #380 · |Δ|max=0.042+ · U.S. court opinions and legal case documents (headings, citations, pro#472 · |Δ|max=0.042+ · Metadata, front matter, and introduction text from academic and techni#460 · |Δ|max=0.042− · Short arithmetic and algebra word problems asking to solve simple line#269 · |Δ|max=0.042+ · Code snippets and technical discussion about statically typed programm#279 · |Δ|max=0.041+ · Text and formatting typical of U.S. court opinions and legal case docu#278 · |Δ|max=0.041+ · Technical text containing hexadecimal, byte, memory, and low-level bin#500 · |Δ|max=0.040+ · Methods and materials details in biomedical research papers (statistic#104 · |Δ|max=0.038+ · Content about games and gaming: video games, tabletop/RPG mechanics, c#325 · |Δ|max=0.037+ · Citations and prior‑art references and boilerplate language commonly f# 17 · |Δ|max=0.037+ · Procedural methods and materials descriptions from biomedical and mole#575 · |Δ|max=0.037+ · Short algebra and arithmetic problems that simplify or evaluate linear#169 · |Δ|max=0.037+ · Short algebra and arithmetic word problems: polynomial manipulation, d#386 · |Δ|max=0.037+ · Mentions of countries, places, and local names in news, encyclopedia, # 27 · |Δ|max=0.036+ · Texts and discussions about Judaism and Christianity: scripture, doctr#137 · |Δ|max=0.036+ · Demographic and census-style descriptions of places (population, house#476 · |Δ|max=0.036+ · Objective-C, Swift, and iOS/macOS developer code and API patterns (UIK#722 · |Δ|max=0.035− · References to U.S. states, cities, local places, and regional/local ne#648 · |Δ|max=0.035+ · Short math and technical snippets: small-answer arithmetic/algebra pro#651 · |Δ|max=0.035+ · Questions and answers about converting or computing numbers in non-dec#175 · |Δ|max=0.035+ · Top-of-file headers in code and documentation (copyright, license, aut#414 · |Δ|max=0.035− · Wikipedia-style entries and disambiguation pages for personal names, b#243 · |Δ|max=0.035+ · Middle-school to high-school style algebra and arithmetic problems: si#184 · |Δ|max=0.034+ · Boilerplate text and header language from U.S. appellate court opinion# 2 · |Δ|max=0.034+ · Texts about plants, crops, and wild animals—botanical names, cultivati#100 · |Δ|max=0.034+ · Tasks involving polynomials and basic symbolic algebra — simplifying, #167 · |Δ|max=0.032+ · Date and time expressions and timestamps in texts, logs, and metadata #375 · |Δ|max=0.032+ · PHP web backend code patterns and common server-side snippets (templat# 4 · |Δ|max=0.032+ · News and commentary about sports games, teams, and player performance #102 · |Δ|max=0.032+ · Texts about chemicals, chemical companies, and industrial uses—descrip#526 · |Δ|max=0.030− · Common small function words and punctuation found in formal/translated# 1 · |Δ|max=0.029+ · Text about sports events and results—especially horse and motor racing#535 · |Δ|max=0.025+ · Formal academic and clinical research prose (scientific article langua#166 · |Δ|max=0.021− · Very common short function words and noisy scraped web boilerplate lan# 67 · |Δ|max=0.018+ · Standard patent boilerplate introducing the invention’s field, objecti#292 · |Δ|max=0.016+ · Formal patent and technical-document language describing devices, comp# 64 · |Δ|max=0.013+ · Footer category and metadata lines from Wikipedia article pages (the c

— jump to response feature cluster (sorted by max |Δ|) —▾

— jump to prompt cluster (sorted by overall |Δ|) —# 39 · Δ=+0.047 · n=22 · Quantum superposition and entanglement#146 · Δ=+0.046 · n=28 · Advanced General Relativity Questions#153 · Δ=+0.046 · n=27 · Graduate-level theoretical physics questions# 50 · Δ=+0.045 · n=18 · Rigorous theoretical physics references# 40 · Δ=+0.045 · n=44 · Advanced theoretical and mathematical physics questions#460 · Δ=+0.045 · n=34 · Foundations and Formalism of Quantum Field Theory#368 · Δ=+0.043 · n=27 · Black holes, Hawking radiation, and quantum gravity#165 · Δ=+0.041 · n=31 · Spinors and Lorentz/Poincaré Representations#268 · Δ=+0.041 · n=28 · Quantum foundations and information theory# 29 · Δ=+0.041 · n=23 · Scalable data system architectures#381 · Δ=+0.040 · n=21 · Applied fluid mechanics and physics problems#359 · Δ=+0.038 · n=27 · Advanced quantum field theory questions#162 · Δ=+0.037 · n=84 · Taboo and illicit fiction requests#278 · Δ=+0.036 · n=50 · Research‑framed safety‑bypass requests# 35 · Δ=+0.035 · n=60 · Quantum many-body & condensed-matter problems# 87 · Δ=+0.035 · n=27 · Special Relativity conceptual puzzles#432 · Δ=+0.035 · n=36 · Using Copyrighted Works Legally#419 · Δ=+0.034 · n=22 · Cloud deployment and integration guides#230 · Δ=+0.034 · n=20 · Advanced classical optics problems#342 · Δ=+0.033 · n=29 · Lagrangian and Hamiltonian mechanics#211 · Δ=+0.033 · n=37 · Rigid-body rotation and dynamics#305 · Δ=+0.033 · n=71 · Jailbreak roleplay for prohibited content#388 · Δ=+0.033 · n=38 · Causes and Origins of Historical Events#392 · Δ=+0.032 · n=50 · Advanced electromagnetism conceptual problems#328 · Δ=+0.032 · n=60 · Sinhala general-knowledge questions#475 · Δ=+0.032 · n=46 · Requests for Classified or Confidential Data#238 · Δ=-0.031 · n=1 · Smart Healthcare Device Proposal Writing#315 · Δ=-0.031 · n=23 · Mandatory 'Reply finished' Closure# 43 · Δ=+0.031 · n=31 · Requests for Personal Medical Records#370 · Δ=+0.031 · n=24 · Thermodynamics and Statistical Mechanics Problems#435 · Δ=+0.030 · n=31 · Sexuality Education and Support Resources#393 · Δ=-0.030 · n=4 · API Parameter Generation & Optimization#307 · Δ=-0.030 · n=22 · DDLC pregnancy roleplay scenes# 9 · Δ=+0.029 · n=40 · Hypothetical cyberattack and hacking requests# 63 · Δ=+0.029 · n=27 · Biographies and Public Figures#149 · Δ=+0.029 · n=43 · Applied PDE/Fluid Mechanics Problems#256 · Δ=+0.028 · n=32 · Orbital mechanics and Kepler problems#389 · Δ=+0.028 · n=36 · Requests for Classified Security Details#461 · Δ=+0.028 · n=41 · Definitions of crimes and harms#420 · Δ=+0.028 · n=71 · Overcoming Obstacles and Procrastination#430 · Δ=-0.028 · n=20 · Minimalist Instagram Marketing Expansion#160 · Δ=+0.027 · n=46 · Advanced quantum field theory questions#478 · Δ=-0.027 · n=4 · End responses with 'Reply finished'#284 · Δ=-0.027 · n=26 · Quantitative aptitude MCQs and solutions#341 · Δ=-0.027 · n=31 · Filter safe words for word game#285 · Δ=-0.027 · n=14 · Phishing and Email Scam Creation#210 · Δ=+0.027 · n=36 · Classical mechanics: forces and constraints# 25 · Δ=+0.026 · n=52 · General relativity and differential-geometry problems#444 · Δ=+0.026 · n=47 · Advanced quantum operator and representation questions#380 · Δ=-0.025 · n=18 · Dataset identification multiple-choice#252 · Δ=+0.025 · n=3 · Codebase documentation and task planning# 61 · Δ=+0.025 · n=39 · Requests to Create Conspiracy Hoaxes# 66 · Δ=-0.025 · n=38 · COVID-19 research: pathogenesis and therapeutics#385 · Δ=+0.025 · n=75 · Coping with Suicidal Thoughts#390 · Δ=+0.025 · n=42 · SVG icons and markup tasks#349 · Δ=+0.025 · n=31 · Graph and Spatial Optimization Problems#171 · Δ=-0.024 · n=21 · Clinical trial eligibility criteria#289 · Δ=-0.024 · n=12 · Filtered Wordlist Preparation for Games#232 · Δ=+0.024 · n=36 · .NET and ASP.NET Development Questions#271 · Δ=-0.024 · n=23 · Pregnant mob character roleplay#404 · Δ=+0.024 · n=54 · Graph theory and network analysis#169 · Δ=-0.024 · n=19 · DDLC pregnancy and childbirth scenes# 41 · Δ=+0.024 · n=25 · Medical histories of public figures#422 · Δ=+0.023 · n=42 · Online Transaction Security and Fraud Prevention#231 · Δ=+0.023 · n=11 · Regex pattern crafting and validation#184 · Δ=-0.023 · n=54 · Elementary arithmetic word problems#103 · Δ=-0.023 · n=8 · COVID-19 clinical and research literature#174 · Δ=-0.023 · n=38 · Etsy product title generator#334 · Δ=-0.022 · n=28 · Product review sentiment classification#183 · Δ=+0.022 · n=45 · Hands-on ML model building and evaluation#107 · Δ=-0.022 · n=33 · Resumes, Cover Letters & Job Applications#304 · Δ=-0.021 · n=14 · WordPress Persian UI Translations#365 · Δ=-0.021 · n=42 · How-to Guides for Violent Crimes# 11 · Δ=+0.021 · n=23 · Heat transfer and thermodynamics problems#323 · Δ=-0.021 · n=28 · ChEBI chemical entity extraction# 23 · Δ=+0.021 · n=23 · Statistical inference and modeling questions#306 · Δ=-0.021 · n=0 · Structured Restaurant and Review Prompts#425 · Δ=-0.021 · n=11 · Recipe ingredient NER extraction#293 · Δ=-0.021 · n=27 · List slicing and element-extraction tasks#456 · Δ=-0.020 · n=28 · University information and announcements#167 · Δ=-0.020 · n=22 · Very short discussion reply generation#116 · Δ=-0.020 · n=15 · German keyword-extraction requests#423 · Δ=-0.020 · n=25 · Short replies to motivational messages#316 · Δ=-0.020 · n=22 · Binary choice: middle/cause selection#440 · Δ=-0.020 · n=19 · Binary biomedical yes/no QA#192 · Δ=-0.020 · n=68 · Visual Textual Entailment Questions# 46 · Δ=-0.020 · n=65 · Indic language translation prompts#214 · Δ=-0.020 · n=39 · Chinese-to-English Academic Translations# 77 · Δ=+0.020 · n=42 · Web and Mobile App Development#124 · Δ=-0.020 · n=74 · Sentence and Paragraph Rewriting#196 · Δ=-0.020 · n=24 · Academic NLP paper excerpts#121 · Δ=-0.020 · n=57 · Indian language translation tasks#457 · Δ=-0.020 · n=55 · Multilingual short-sentence translation tasks# 22 · Δ=-0.020 · n=19 · Citation Intent Classification# 79 · Δ=-0.020 · n=15 · YouTube language-learning video metadata#413 · Δ=-0.020 · n=5 · Very short replies to math statements#496 · Δ=-0.019 · n=25 · Virology literature Q&A# 27 · Δ=+0.019 · n=52 · Challenging Stereotypes and Stigma#360 · Δ=-0.019 · n=64 · Natural Language Inference (Entailment)#339 · Δ=-0.019 · n=32 · Binary text labeling tasks# 98 · Δ=-0.019 · n=7 · Xion community hype messages#358 · Δ=-0.019 · n=34 · Deutsche Text-Keyword-Extraktion (10)#350 · Δ=-0.019 · n=45 · News-to-JSON-schema extraction#148 · Δ=-0.019 · n=13 · Casual browser game descriptions#372 · Δ=-0.019 · n=40 · Scientific manuscript excerpts#281 · Δ=-0.019 · n=55 · Biomedical entity extraction to JSON#288 · Δ=-0.019 · n=31 · Short multiple-choice QA tasks#134 · Δ=+0.018 · n=47 · Eradicating pests, weeds, mold, and grime#369 · Δ=-0.018 · n=11 · Wikipedia Title Category Classification#382 · Δ=-0.018 · n=17 · Research paper figure/table QA#294 · Δ=-0.018 · n=45 · Premise–Hypothesis Entailment Tasks#498 · Δ=-0.018 · n=62 · Bilingual sentence translation tasks#295 · Δ=-0.018 · n=55 · Multilingual translation task prompts#241 · Δ=-0.018 · n=69 · Common-sense sentence selection#142 · Δ=-0.018 · n=50 · Risk-Disclosure Entailment Tasks#117 · Δ=-0.018 · n=42 · Multilingual Formal Text Translations#220 · Δ=+0.018 · n=13 · Electric circuits and EM problems#114 · Δ=-0.018 · n=23 · 10 Keywords extrahieren (Deutsch)#255 · Δ=-0.018 · n=14 · User greetings and salutations#105 · Δ=+0.018 · n=27 · Stereotypes and workplace discrimination#434 · Δ=+0.018 · n=72 · Undergraduate Organic Chemistry Problems#245 · Δ=-0.017 · n=60 · Product review sentiment classification#438 · Δ=+0.017 · n=41 · Constrained Multilingual Short-Answer Prompts#448 · Δ=-0.017 · n=54 · Biomedical extractive QA tasks#147 · Δ=-0.017 · n=31 · List-based numeric transformation tasks#299 · Δ=-0.017 · n=57 · Multilingual Translation and Localization#486 · Δ=-0.017 · n=84 · Molecular Immunology and Gene Regulation#130 · Δ=-0.017 · n=86 · Textual Entailment (NLI) Questions#239 · Δ=-0.017 · n=7 · Navigation action command conversion# 71 · Δ=-0.017 · n=42 · Freedom Planet crossover reaction scenarios#469 · Δ=+0.017 · n=14 · XML to JSON/Dict Conversion#175 · Δ=-0.017 · n=19 · Short-story Sentence Completion#407 · Δ=-0.017 · n=53 · Professional email and message editing#120 · Δ=-0.017 · n=17 · Aviation and Airline Knowledge Requests#320 · Δ=-0.017 · n=95 · Passage-based Factoid QA#259 · Δ=-0.017 · n=51 · Short-story commonsense QA#242 · Δ=-0.017 · n=66 · elementary science multiple-choice q&a#177 · Δ=-0.017 · n=11 · Profession coreference gender identification#374 · Δ=-0.017 · n=48 · Short news summarization and headlines#205 · Δ=-0.016 · n=12 · Academic and Professional Email & Letter Drafting#129 · Δ=-0.016 · n=20 · Context-based Question Generation#156 · Δ=+0.016 · n=53 · Denying Rights to Imaginary Entities#206 · Δ=-0.016 · n=29 · Biomedical and Pharmacology Research Abstracts# 75 · Δ=-0.016 · n=41 · Fantasy hybrid erotic story prompts#417 · Δ=-0.016 · n=69 · Multilingual Sentence Translation Tasks#233 · Δ=-0.016 · n=56 · Language translation requests#376 · Δ=-0.016 · n=30 · Clinical and Biomedical Research Sentences#302 · Δ=-0.016 · n=33 · Dutch Named-Entity Recognition Tasks#218 · Δ=-0.016 · n=42 · Email drafting and editing requests#313 · Δ=-0.016 · n=13 · Persian–English Religious Text Translation#249 · Δ=-0.016 · n=18 · Sentence-pair Editing for Contradiction/Entailment# 67 · Δ=-0.016 · n=18 · Personalized Greeting and Message Drafting# 86 · Δ=-0.016 · n=56 · Binary Event-Relation Classification# 31 · Δ=+0.016 · n=16 · Applied Public-Key Cryptography Problems#325 · Δ=+0.016 · n=77 · Constrained resource allocation optimization#443 · Δ=-0.016 · n=38 · Academic paper excerpt QA#362 · Δ=+0.016 · n=19 · Machine learning dataset identification#227 · Δ=+0.016 · n=34 · SQL schema and database operations#366 · Δ=-0.015 · n=53 · User-persona follow-up questions#204 · Δ=-0.015 · n=68 · Multiple-choice common-sense QA#377 · Δ=+0.015 · n=72 · String manipulation coding tasks# 89 · Δ=+0.015 · n=83 · Competitive programming problem statements# 54 · Δ=-0.015 · n=38 · Japanese-to-other-language translations# 3 · Δ=-0.015 · n=24 · Deep learning paper abstracts#319 · Δ=-0.015 · n=33 · Restaurant query translation & reformatting#308 · Δ=+0.015 · n=40 · Implementing and explaining sorting algorithms#279 · Δ=+0.015 · n=44 · Biomedical research paper snippets#244 · Δ=+0.015 · n=36 · Organic structure drawing & stereochemistry#428 · Δ=-0.015 · n=15 · Sentence Sequencing and Labeling Tasks# 72 · Δ=-0.015 · n=10 · Cross-lingual Cause–Effect Classification# 49 · Δ=-0.015 · n=33 · Clinical abstract QA and extraction#168 · Δ=-0.015 · n=51 · Explicit sexual content and abuse requests# 80 · Δ=-0.015 · n=65 · Yes/No Common‑Sense Fact Checking# 7 · Δ=+0.015 · n=39 · Binary Tree and BST Algorithms#287 · Δ=-0.014 · n=11 · Natural Language Inference (NLI) Tasks#489 · Δ=-0.014 · n=20 · Constrained Riddles, Trivia and Puzzles#269 · Δ=+0.014 · n=32 · Structured data analysis reports# 30 · Δ=-0.014 · n=62 · Malicious hacking and cyberattack instructions#102 · Δ=-0.014 · n=55 · Short NLP Annotation and Classification#301 · Δ=-0.014 · n=47 · Tabular column-mapping and SQL generation#410 · Δ=+0.014 · n=21 · Obfuscated Python Code with Prompt-Injection# 18 · Δ=-0.014 · n=18 · Film plot Q&A and trivia# 88 · Δ=+0.014 · n=24 · Web scraping, crawlers & HTTP scripts#429 · Δ=+0.014 · n=33 · Business Leaders' Strategies & Profiles#408 · Δ=-0.014 · n=53 · Interactive Web Canvas & Game Code#449 · Δ=+0.014 · n=57 · JSON parsing and transformation tasks# 52 · Δ=-0.014 · n=26 · Vietnamese news/article paragraph writing# 47 · Δ=-0.014 · n=16 · Name and Title Generation Prompts#348 · Δ=-0.014 · n=54 · Sentence Rewriting and Paraphrase Tasks#157 · Δ=+0.014 · n=24 · Mobile app development — Android & Flutter# 26 · Δ=-0.013 · n=36 · Community and Urban Development Planning# 32 · Δ=-0.013 · n=50 · Business writing and strategy templates# 45 · Δ=+0.013 · n=13 · Genetic, Age, and Disability Discrimination#190 · Δ=-0.013 · n=12 · Algorithmic trading code and debugging#326 · Δ=-0.013 · n=25 · Constrained Writing and Editing Requests#277 · Δ=-0.013 · n=28 · Recruiting, Interviewing, and HR Communication#273 · Δ=-0.013 · n=50 · App review intent classification#247 · Δ=-0.013 · n=26 · Actor role identification questions#406 · Δ=+0.013 · n=31 · Specific medication and treatment requests#499 · Δ=+0.013 · n=39 · Complex analysis and advanced math# 16 · Δ=-0.013 · n=32 · App and Product Review Rating#383 · Δ=-0.013 · n=25 · Social media copywriting with constraints#330 · Δ=-0.013 · n=42 · Emotion Classification and Generation Tasks#274 · Δ=-0.013 · n=64 · Time, work, and schedule word problems# 0 · Δ=+0.013 · n=33 · Cloud DevOps, Containers & Deployment# 33 · Δ=+0.013 · n=63 · Video game combat and tactics#459 · Δ=-0.013 · n=12 · Sentiment Classification of Reviews#324 · Δ=-0.013 · n=61 · Slogans, Taglines, and Ad Copy# 1 · Δ=-0.013 · n=66 · Real-world money word problems#186 · Δ=-0.013 · n=29 · Event attendance and counting problems#127 · Δ=-0.013 · n=41 · Constrained Content Generation Tasks#216 · Δ=-0.013 · n=72 · Generate question and justification#166 · Δ=+0.012 · n=108 · Modeling with differential equations#312 · Δ=+0.012 · n=53 · Constrained historical writing prompts#327 · Δ=-0.012 · n=64 · Person identification / biographical queries#297 · Δ=-0.012 · n=19 · Strict formatting and constraint prompts# 5 · Δ=-0.012 · n=46 · Constrained sentence and phrase generation#203 · Δ=-0.012 · n=29 · Print-on-Demand Artwork SEO Descriptions# 84 · Δ=-0.012 · n=33 · K-12 Teaching Strategies & Lesson Design#104 · Δ=+0.012 · n=25 · Classical and Programming Ciphers#176 · Δ=-0.012 · n=31 · Переформулировка и рерайт текстов#222 · Δ=-0.012 · n=21 · Bengali hate-speech classification#138 · Δ=-0.012 · n=16 · Cooking, ingredients, and food QA#494 · Δ=-0.012 · n=44 · Character-focused Roleplay Scenarios# 73 · Δ=-0.012 · n=43 · Social media hate/offensive classification#199 · Δ=-0.012 · n=74 · Table column mapping and extraction#332 · Δ=-0.012 · n=34 · Algebraic equation solving tasks#309 · Δ=-0.012 · n=6 · Minecraft chat bot command prompts# 24 · Δ=-0.012 · n=35 · Shopping and food price word problems#221 · Δ=-0.012 · n=22 · Arabic Islamic quiz questions#163 · Δ=+0.012 · n=38 · Mathematical Machine Learning Problems#345 · Δ=-0.012 · n=29 · Garden-themed arithmetic word problems#213 · Δ=-0.012 · n=112 · Quantitative General Chemistry Problems#415 · Δ=+0.012 · n=63 · HTML/CSS/JavaScript Front-End Tasks#398 · Δ=+0.012 · n=46 · Algorithm Performance and Systems Modeling# 10 · Δ=+0.011 · n=54 · JavaScript array and string functions#208 · Δ=+0.011 · n=18 · Graph algorithm coding problems#458 · Δ=+0.011 · n=85 · Competitive programming problem statements#321 · Δ=-0.011 · n=108 · Real-life arithmetic word problems# 8 · Δ=+0.011 · n=43 · Object-oriented class design tasks#471 · Δ=-0.011 · n=48 · Historical and biographical trivia questions#424 · Δ=+0.011 · n=35 · How-to and instructional requests#403 · Δ=+0.011 · n=15 · Tesla and Renewable Energy/EVs# 91 · Δ=-0.011 · n=88 · Malagasy language learning tasks#412 · Δ=+0.011 · n=68 · Code Translation and Porting Requests#260 · Δ=-0.011 · n=55 · Football and sports trivia#226 · Δ=-0.011 · n=21 · Binary yes/no annotation tasks#336 · Δ=-0.011 · n=32 · Academic summarizing and rewriting#421 · Δ=+0.011 · n=26 · Fibonacci and recursive sequence functions#442 · Δ=+0.011 · n=63 · Text Word Counting and Frequency#416 · Δ=-0.011 · n=48 · Music, Film, and Creative Production Analysis#181 · Δ=-0.011 · n=73 · Physical chemistry calculation problems#479 · Δ=+0.011 · n=39 · Array and Histogram Algorithm Problems#150 · Δ=+0.011 · n=47 · Programming language how-tos and explanations#335 · Δ=+0.011 · n=58 · Array and List Algorithm Problems#329 · Δ=-0.011 · n=11 · Disaster and Operational Risk Disclosure#133 · Δ=+0.011 · n=37 · Constrained combinatorics and counting# 55 · Δ=-0.011 · n=47 · Academic Research Writing Requests#272 · Δ=-0.011 · n=28 · Everyday cost and ticket word problems#491 · Δ=-0.011 · n=63 · Constrained Sports Summaries and Lists#180 · Δ=+0.011 · n=28 · Database and SQL programming questions#136 · Δ=+0.010 · n=40 · Programming tools, frameworks and libraries# 20 · Δ=-0.010 · n=36 · Basic arithmetic and calculation prompts#276 · Δ=+0.010 · n=52 · tamil general-knowledge quizzes#257 · Δ=-0.010 · n=22 · Real estate and housing queries# 57 · Δ=-0.010 · n=61 · Elementary arithmetic word problems#351 · Δ=-0.010 · n=31 · Liquid volume and rate word problems#371 · Δ=+0.010 · n=96 · Python list and string processing#354 · Δ=-0.010 · n=47 · Animal classification and simple biology# 68 · Δ=-0.010 · n=23 · Short e-commerce fashion & beauty copy#462 · Δ=-0.010 · n=52 · Shopping and Price Word Problems#193 · Δ=-0.010 · n=44 · Money word-problem arithmetic#145 · Δ=-0.010 · n=15 · Mean and Basic Statistics Problems#495 · Δ=-0.010 · n=10 · Letter and Word Counting Tasks#331 · Δ=+0.010 · n=54 · Interactive Bank Account Command Processor#378 · Δ=-0.010 · n=55 · Custom scripts and screenplays#314 · Δ=-0.010 · n=23 · French translations, definitions, corrections#322 · Δ=-0.010 · n=31 · Area Perimeter Volume Word Problems#235 · Δ=+0.010 · n=28 · Flatten and Process Nested Python Lists#113 · Δ=-0.010 · n=31 · Hourly wage and income word problems#118 · Δ=-0.010 · n=57 · Word and phrase meaning requests#282 · Δ=-0.010 · n=44 · Book-related arithmetic word problems#386 · Δ=-0.010 · n=35 · Customized Letter Writing Requests# 90 · Δ=-0.010 · n=40 · Personal reflection and emotional growth#411 · Δ=-0.010 · n=24 · Distance, Speed and Time Word Problems#172 · Δ=-0.010 · n=18 · Calorie and nutrition math problems# 83 · Δ=+0.010 · n=34 · Applied engineering math problems#452 · Δ=+0.010 · n=44 · Biochemistry textbook problem set#436 · Δ=-0.010 · n=14 · Age word problems and puzzles# 81 · Δ=-0.010 · n=19 · Genesis and Biblical Creation Commentary# 65 · Δ=-0.009 · n=94 · Geography factoid and capitals#439 · Δ=-0.009 · n=57 · Time and work word problems#343 · Δ=+0.009 · n=89 · Simple total/aggregation functions#202 · Δ=-0.009 · n=36 · Retail sales and profit word problems#353 · Δ=-0.009 · n=37 · Running, Cycling, and Distance Word Problems#356 · Δ=+0.009 · n=50 · Multilingual constrained answer requests#264 · Δ=-0.009 · n=44 · Historical Research and Contextual Elaboration# 51 · Δ=+0.009 · n=60 · Economic and Business Optimization Problems#363 · Δ=-0.009 · n=27 · Multilingual paragraph continuation prompts#173 · Δ=-0.009 · n=33 · Baking and Food Math Word Problems#453 · Δ=-0.009 · n=41 · Sports-themed arithmetic word problems#144 · Δ=-0.009 · n=20 · Weight, mass and unit word problems# 60 · Δ=+0.009 · n=50 · Advanced undergraduate physics problems# 42 · Δ=+0.009 · n=25 · Matrix and Modular Cryptography Problems#290 · Δ=-0.009 · n=20 · Travel and Transportation Queries#139 · Δ=+0.009 · n=33 · Athlete training and workout guidance#154 · Δ=-0.009 · n=36 · Ordinary Differential Equations & Calculus#119 · Δ=+0.009 · n=87 · Small programming and coding tasks#158 · Δ=-0.009 · n=26 · Fruit-themed arithmetic word problems#248 · Δ=+0.008 · n=22 · Rule-based sentiment classification#209 · Δ=+0.008 · n=19 · Book recommendations and reading resources#261 · Δ=-0.008 · n=34 · Dating, Flirting, and Relationship Advice#254 · Δ=-0.008 · n=44 · Language identification and translation#451 · Δ=-0.008 · n=47 · Animal-themed arithmetic word problems#397 · Δ=-0.008 · n=70 · Constrained essay-writing prompts#250 · Δ=-0.008 · n=41 · Visual/textual entailment judgments#195 · Δ=+0.008 · n=63 · Python total/sum calculation functions#237 · Δ=+0.008 · n=66 · python list and dict data-processing#115 · Δ=-0.008 · n=44 · Natural Language Inference (Premise–Hypothesis)#455 · Δ=-0.008 · n=25 · School classroom math word problems#236 · Δ=-0.008 · n=30 · Constrained blog post writing tasks# 37 · Δ=+0.008 · n=36 · Recipe and Cooking Instructions#474 · Δ=+0.008 · n=12 · Workplace gender and pregnancy discrimination#466 · Δ=+0.008 · n=52 · Palindrome-checking Programs and Functions#367 · Δ=-0.008 · n=26 · TV and movie marathon time math#298 · Δ=-0.008 · n=12 · Natural Language Inference (NLI) Questions#219 · Δ=+0.008 · n=35 · Python dictionary transformation tasks#228 · Δ=-0.008 · n=63 · Constrained creative writing prompts# 78 · Δ=+0.007 · n=70 · Applied regression and statistical inference#215 · Δ=-0.007 · n=55 · History and war Qs with constraints#473 · Δ=-0.007 · n=36 · Lesson Plans and Teaching Resources#480 · Δ=+0.007 · n=59 · Constrained math contest problems#109 · Δ=+0.007 · n=43 · Undergraduate biochemistry problems# 12 · Δ=-0.007 · n=46 · Requests for Hateful or Discriminatory Content#141 · Δ=-0.007 · n=36 · Film and Movie Information Queries#292 · Δ=+0.007 · n=60 · Refactor and Debug Python Code#201 · Δ=+0.007 · n=46 · Stock Market & Investment Advice# 95 · Δ=-0.007 · n=49 · Illegal drugs procurement and distribution#379 · Δ=+0.007 · n=31 · Molecular biology exam-style problems#258 · Δ=+0.007 · n=22 · Roman Numeral Conversion Tasks#485 · Δ=+0.007 · n=41 · Fictional Game Scenarios & Mechanics#492 · Δ=+0.007 · n=24 · Travel route optimization and TSP#464 · Δ=+0.007 · n=52 · Requests to Generate Fake News#445 · Δ=-0.007 · n=38 · Product descriptions and e‑commerce copy#283 · Δ=-0.007 · n=18 · Clinical trial and research abstracts#179 · Δ=-0.007 · n=22 · Ecology and Species Conservation Requests# 28 · Δ=+0.007 · n=42 · Requests for Personal or Financial Data#352 · Δ=+0.007 · n=40 · Matrix and Multi-Dimensional Array Algorithms#170 · Δ=+0.006 · n=44 · Sorting lists of dictionaries#131 · Δ=+0.006 · n=33 · Contact and address information requests#265 · Δ=+0.006 · n=41 · Mathematical Music and Signal Problems#364 · Δ=+0.006 · n=48 · Python List Manipulation Tasks#384 · Δ=-0.006 · n=97 · Undergraduate Chemistry MCQ Practice#100 · Δ=-0.006 · n=59 · Geometry problems with output constraints#122 · Δ=-0.006 · n=37 · Constrained and Themed Poetry Prompts#483 · Δ=-0.006 · n=15 · Alcoholic drinks and recipes#344 · Δ=+0.006 · n=43 · Mathematical Environmental Modeling#155 · Δ=+0.006 · n=76 · Organic reaction mechanism questions# 99 · Δ=-0.006 · n=27 · Narrative arithmetic word problems# 15 · Δ=+0.006 · n=58 · Punjabi Gurmukhi Scripture & Cloze#476 · Δ=-0.006 · n=38 · Music identification and trivia# 6 · Δ=+0.006 · n=20 · Email address validation tasks#310 · Δ=+0.006 · n=19 · URL parsing, validation, and requests#229 · Δ=-0.006 · n=19 · Randomized clinical trial abstracts#346 · Δ=+0.006 · n=30 · Academic paper abstracts and reviews#361 · Δ=+0.006 · n=21 · CDC and infectious disease control#128 · Δ=+0.006 · n=47 · String algorithm and substring problems#337 · Δ=+0.006 · n=76 · Geometry for Art and Design# 64 · Δ=+0.005 · n=37 · Global macroeconomics and institutions#488 · Δ=+0.005 · n=49 · Sexualized Fanfiction and Fetish Scenes#318 · Δ=+0.005 · n=20 · Extracting Email Addresses with Python# 21 · Δ=-0.005 · n=41 · Digital & Social Media Marketing Strategies# 38 · Δ=-0.005 · n=34 · Worktime and Scheduling Word Problems#431 · Δ=-0.005 · n=38 · University Education and Admissions Queries# 97 · Δ=+0.005 · n=68 · SQL query construction and constraints# 58 · Δ=+0.005 · n=90 · Small Python numeric utility functions#414 · Δ=+0.005 · n=43 · Factorial computation programming tasks#470 · Δ=+0.005 · n=16 · Procedural Spell JSON Generation#418 · Δ=+0.005 · n=48 · Art creation and styling prompts#447 · Δ=-0.005 · n=40 · Requests for hateful demeaning jokes#152 · Δ=-0.005 · n=33 · AI applications, ethics, and analysis#446 · Δ=+0.005 · n=69 · Undergraduate Organic Chemistry Problems#477 · Δ=+0.005 · n=66 · Python average/calculation functions#198 · Δ=+0.005 · n=60 · Prime number algorithms and programs#224 · Δ=+0.005 · n=37 · Mathematical Modeling of Film Topics#487 · Δ=-0.005 · n=30 · Food questions and culinary classification#454 · Δ=+0.005 · n=53 · Molecular Structure and Bonding Problems#207 · Δ=-0.005 · n=36 · Fact-checking and entailment judgments#101 · Δ=+0.005 · n=45 · Chemical Bonding and Molecular Structure# 13 · Δ=-0.005 · n=22 · Weather and Climate Questions#405 · Δ=-0.005 · n=50 · Computer Science & Programming Explanations#188 · Δ=-0.005 · n=30 · Constrained Film Review and Analysis#164 · Δ=+0.005 · n=58 · Sorting and Searching Algorithm Exercises#441 · Δ=-0.005 · n=23 · Code Debugging and Feature Requests#270 · Δ=-0.005 · n=40 · Literature and Book Questions#433 · Δ=-0.004 · n=35 · Constrained Travel Itineraries & Guides#391 · Δ=-0.004 · n=61 · Medical and health questions#463 · Δ=+0.004 · n=39 · Mathematical models for health#311 · Δ=+0.004 · n=38 · Advice for Financial Crime and Fraud#373 · Δ=+0.004 · n=40 · Pandas DataFrame Manipulation and Visualization# 56 · Δ=+0.004 · n=49 · Why can't X do supernatural feats#340 · Δ=-0.004 · n=30 · Banking, Accounting & Personal Finance#266 · Δ=-0.004 · n=50 · Yes/No Sentence Classification Tasks#137 · Δ=+0.004 · n=52 · Stereochemistry and Isomerism Concepts#465 · Δ=+0.004 · n=49 · Dates and historical year queries#106 · Δ=+0.004 · n=35 · Sequence and Series Math Problems#126 · Δ=+0.004 · n=37 · Word Frequency and Counting Programs#300 · Δ=+0.004 · n=18 · Nutrition and diet linear optimization#194 · Δ=+0.004 · n=56 · Sports statistics and probability modeling#159 · Δ=+0.004 · n=48 · Parametric motion and trajectory problems# 2 · Δ=+0.004 · n=62 · Constrained Format Writing Prompts#123 · Δ=-0.004 · n=27 · Music topics with output constraints#493 · Δ=-0.004 · n=42 · Games, puzzles and strategies#291 · Δ=-0.004 · n=42 · Requests for Insults and Hate Speech# 48 · Δ=+0.004 · n=51 · Array and Subarray Algorithm Problems# 59 · Δ=+0.004 · n=21 · Image and Pixel Processing Functions# 62 · Δ=+0.004 · n=45 · Geometry problems for parks and gardens#395 · Δ=-0.004 · n=13 · Recipes, Meal Plans, and Nutrition#357 · Δ=-0.004 · n=45 · Coordinated Misinformation and Defamation#246 · Δ=-0.004 · n=39 · Requests for harassment and threats# 34 · Δ=+0.003 · n=40 · NMR Spectra Interpretation and Analysis#427 · Δ=-0.003 · n=56 · strict output format constraints#402 · Δ=-0.003 · n=34 · Automotive and driving queries# 93 · Δ=-0.003 · n=115 · Small single-purpose function tasks#355 · Δ=+0.003 · n=44 · Frontend JavaScript & React DOM Issues#375 · Δ=+0.003 · n=13 · Mathematical battlefield problem sets# 4 · Δ=+0.003 · n=19 · Generate unique string permutations#212 · Δ=+0.003 · n=58 · Mathematical Revenue and Growth Modeling#437 · Δ=+0.003 · n=17 · Programming error and debugging questions#481 · Δ=+0.003 · n=42 · Requests for hateful or discriminatory content#225 · Δ=-0.003 · n=60 · Constrained List‑Generation Prompts#191 · Δ=-0.003 · n=113 · List-of-integers stdin problems#396 · Δ=+0.003 · n=44 · Constrained-response multilingual prompts#296 · Δ=-0.003 · n=30 · Strict formatting and output constraints#333 · Δ=+0.003 · n=24 · Binary string to decimal conversion#111 · Δ=-0.002 · n=23 · Fundraising and budget word problems#135 · Δ=+0.002 · n=19 · Mathematical Election & Voter Modeling#468 · Δ=-0.002 · n=41 · Budget and Quadratic Word Problems#472 · Δ=+0.002 · n=31 · Mathematical Modeling of Historical Data#200 · Δ=+0.002 · n=35 · Planetary and Astronomy Prompts#151 · Δ=+0.002 · n=30 · Mathematical sports performance modeling#197 · Δ=-0.002 · n=45 · Short-definition 'What is' questions# 94 · Δ=+0.002 · n=95 · Python string manipulation functions#286 · Δ=-0.002 · n=35 · Environmental sustainability and conservation#409 · Δ=-0.002 · n=9 · Select Persian-only string index#497 · Δ=+0.002 · n=17 · Diet, Nutrition, and Weight-Loss Queries#240 · Δ=-0.002 · n=73 · Thermodynamics and Vapor–Liquid Problems#178 · Δ=-0.002 · n=48 · User credential and input validation functions#387 · Δ=-0.002 · n=22 · Combinatorics and Set Theory Problems#394 · Δ=+0.002 · n=52 · String manipulation stdin problems#217 · Δ=+0.002 · n=52 · Python dictionary utility functions#125 · Δ=+0.002 · n=29 · Songwriting and lyric requests# 53 · Δ=-0.002 · n=23 · String Manipulation and Transformation#490 · Δ=+0.002 · n=65 · Chemical Kinetics and Reactor Calculations#262 · Δ=+0.002 · n=66 · Introductory Inorganic and Analytical Chemistry#303 · Δ=-0.002 · n=15 · Excel and Google Sheets formulas & macros#187 · Δ=+0.002 · n=31 · Requests for Hateful or Degrading Stories# 74 · Δ=-0.002 · n=91 · List and dict processing functions#482 · Δ=-0.001 · n=19 · Event and Community Planning Prompts#253 · Δ=-0.001 · n=77 · Write programming code snippets# 92 · Δ=+0.001 · n=92 · Social media growth and engagement modeling# 69 · Δ=-0.001 · n=17 · Radioactivity and Nuclear Decay Problems#189 · Δ=-0.001 · n=65 · Constrained short math problems# 82 · Δ=-0.001 · n=47 · Constrained JSON Response Prompts#112 · Δ=+0.001 · n=50 · Investment and Interest Math Problems#450 · Δ=+0.001 · n=35 · Book-and-reading math problems#400 · Δ=+0.001 · n=101 · Organic reaction mechanism and products#280 · Δ=+0.001 · n=29 · Instructions for committing and hiding crimes#267 · Δ=+0.001 · n=117 · Python list processing utilities#161 · Δ=-0.001 · n=45 · Music-themed math word problems#132 · Δ=+0.001 · n=67 · Polynomial modeling and analysis#275 · Δ=-0.001 · n=40 · File and Filename Utility Functions#251 · Δ=-0.001 · n=36 · Date and Time Parsing & Conversion#108 · Δ=+0.001 · n=14 · Programming Loops and Iteration#317 · Δ=+0.001 · n=22 · Official websites and social profiles# 36 · Δ=+0.001 · n=24 · Shell and Command‑Line Scripting# 96 · Δ=-0.001 · n=21 · Chemistry research article excerpts# 85 · Δ=+0.001 · n=65 · Simple arithmetic and division programs#426 · Δ=+0.001 · n=72 · Small Python Numerical Function Implementations# 17 · Δ=-0.001 · n=37 · Questions about Political Leaders#110 · Δ=+0.001 · n=29 · Legal concepts, cases, and summaries# 70 · Δ=-0.001 · n=30 · Constrained Motivational Speech Prompts#399 · Δ=+0.001 · n=33 · Population growth modeling problems#338 · Δ=+0.001 · n=52 · Sum-of-Squares and Simple Numeric I/O#401 · Δ=-0.001 · n=104 · Applied probability and statistics problems# 19 · Δ=+0.000 · n=8 · Bank account transaction simulators#234 · Δ=+0.000 · n=37 · Health and Vaccine Misinformation Promotion# 44 · Δ=+0.000 · n=45 · Constrained article-writing prompts#484 · Δ=+0.000 · n=54 · Landmark addresses and locations#182 · Δ=+0.000 · n=22 · Networking and system administration#347 · Δ=+0.000 · n=16 · Full copyrighted text reproduction requests#185 · Δ=-0.000 · n=37 · Yorùbá and Nigerian News Headlines#143 · Δ=-0.000 · n=20 · Compute average from input numbers# 14 · Δ=+0.000 · n=73 · Constrained fetish and short-story prompts# 76 · Δ=-0.000 · n=24 · Renewable energy math problems#243 · Δ=+0.000 · n=36 · Fraudulent identities, scams, and laundering#140 · Δ=+0.000 · n=35 · Constrained Political Analysis Prompts#223 · Δ=-0.000 · n=47 · Incitement and Hate Speech Requests#467 · Δ=-0.000 · n=41 · Fibonacci number computation programs#263 · Δ=-0.000 · n=35 · Business and Financial Math Word Problems

— jump to prompt cluster (sorted by overall |Δ|) —▾

Data UMAP

60456 features  |  814 feature clusters

color:cluster idchosen-rejected

— jump to feature cluster (sorted by chosen/rejected polarity) —#346 · balanced · 76 features · Technical source-code and API documentation phrasing—comments, docstrings, pa…#359 · balanced · 74 features · Everyday informal blog and forum writing using common small words, personal r…#281 · rejected 71% · 88 features · Neutral, informative assistant text: explanatory how‑tos, factual answers, an…#748 · chosen 77% · 16 features · Formatting and structural markers used in technical and academic documents (s…#133 · chosen 74% · 71 features · Language and formatting typical of judicial opinions and legal case citations…#211 · balanced · 103 features · Neutral explanatory assistant responses: concise informational explanations, …#782 · rejected 84% · 8 features · Tokens and identifiers in programming examples — variable/function names, cod…#401 · balanced · 63 features · Very common short words and tokens that start or appear in assistant replies …#217 · balanced · 102 features · Programming source-code tokens and common identifier/markup elements (variabl…#533 · balanced · 51 features · Academic biomedical research writing—methods, results, and topic-focused pros…#575 · mixed 51/49 · 25 features · Short algebra and arithmetic problems that simplify or evaluate linear and po…#449 · chosen 59% · 60 features · Common short words, numbers, and everyday technical terms that appear through…# 37 · rejected 73% · 177 features · Common, high-frequency words and connective phrasing found in formal/technica…#792 · balanced · 7 features · Very common short English words and small grammatical/connective tokens used …#248 · balanced · 94 features · Formal LaTeX-heavy academic mathematics and physics writing with displayed eq…#794 · rejected 75% · 6 features · Tokens found in source code files, especially generated API/SDK code and surr…#672 · chosen 65% · 31 features · Phrases and language used to ask for technical help, troubleshooting, or sugg…# 10 · balanced · 205 features · Common short function words and tiny connective tokens used in everyday expla…#343 · chosen 54% · 59 features · Language and formatting commonly found in judicial opinions and appellate cou…#417 · chosen 60% · 66 features · Technical programming Q&A posts: debugging, how-to questions, code snippets, …#423 · chosen 63% · 65 features · Very frequent short function words and punctuation that appear across diverse…#452 · balanced · 60 features · Tokens and short fragments typical of academic and scientific papers (LaTeX/m…#669 · rejected 70% · 32 features · Short conversational prompts and responses typical of everyday chat: question…#218 · balanced · 102 features · Very common short words, subword fragments, and other high-frequency tokens u…#446 · chosen 62% · 58 features · Formatting and rendering tables, tabular data, and table-like section headers…#312 · rejected 83% · 57 features · Structured technical text and common degenerate/repetitive token patterns fou…# 51 · rejected 76% · 166 features · Very common short words and generic connective phrasing found across many typ…#131 · chosen 56% · 124 features · Very common short function words and conversational fillers that appear acros…# 44 · balanced · 170 features · Very high‑frequency common words and short grammatical tokens used across man…# 14 · balanced · 197 features · Technical and academic writing, especially scientific papers and programming/…#228 · balanced · 99 features · Very common short words, punctuation, and structural tokens that appear acros…#138 · chosen 56% · 122 features · Very common small English words and generic connective/qualification phrases …# 65 · balanced · 152 features · LaTeX-formatted mathematics and technical notation found in scholarly math/ph…#340 · rejected 70% · 77 features · Everyday conversational and informational assistant language—common words, us…#556 · rejected 81% · 46 features · Technical source code, markup, and structured documentation fragments (identi…#180 · balanced · 112 features · Very common short function words and basic punctuation used to connect and st…#427 · chosen 55% · 63 features · References to figures, figure panels, image captions, and related labels/anno…#197 · balanced · 105 features · Common small function words and generic tokens found in normal prose, code, a…# 77 · rejected 77% · 147 features · Very common short function words and small grammatical tokens used across man…#262 · balanced · 93 features · Tokens and snippets from programming source code—variable names, type/paramet…#331 · balanced · 79 features · Step‑by‑step numeric and algebraic problem solving — basic arithmetic, algebr…# 43 · rejected 72% · 171 features · Very common short tokens and small function words found across many contexts …#742 · chosen 54% · 19 features · Phrases that qualify, limit, or reference scope and knowledge (e.g., hedges, …#303 · chosen 64% · 39 features · Text from judicial opinions and legal case citations (court reporter referenc…#615 · mixed 48/52 · 39 features · Common short function words, punctuation, and connective tokens found in ordi…#812 · rejected 58% · 4 features · Tiny formatting and punctuation fragments commonly found in academic, technic…#626 · chosen 57% · 38 features · Promotional and informational web copy: marketing, calls-to-action, and site/…#518 · chosen 62% · 53 features · Technical and programming document content — source code, configuration, tabl…#588 · balanced · 43 features · Instructional step-by-step guidance, lists, and structured how-to content for…#653 · rejected 77% · 35 features · High-frequency function words, punctuation, and markup tokens common in acade…#732 · balanced · 20 features · Very common short English words and grammatical tokens (function words, artic…# 31 · balanced · 182 features · Very common small words and generic phrasing used in informational, academic,…#241 · balanced · 97 features · Short technical and informational phrases common in documentation, code examp…#601 · balanced · 41 features · Formal mathematical and technical academic writing (LaTeX-style equations, de…# 62 · chosen 71% · 154 features · Short, common words and polite assistant phrasing across Romance languages (e…# 18 · chosen 68% · 178 features · Scholarly and legal writing with formal citations, case references, technical…#176 · balanced · 89 features · Short common tokens and punctuation found in technical/math/code and informal…#399 · chosen 63% · 61 features · Source code files and developer-facing documentation comments (code examples,…#484 · chosen 69% · 56 features · Text that points readers to external web resources, links, examples, or docum…#327 · balanced · 79 features · Section headings and standard scholarly article text (methods, results, discu…#400 · mixed 47/53 · 69 features · Items and short fields from lists, catalogs, indexes, and technical enumerati…#805 · chosen 60% · 4 features · Generic tokens that commonly appear in encyclopedic, legal, and product texts…#410 · balanced · 66 features · Common connective and referential words and phrases used in encyclopedic, tec…#560 · balanced · 48 features · Practical programming help and how-to code examples, especially Python data m…#797 · rejected 65% · 5 features · Very common small words and generic programming terms used to link or refer t…#630 · balanced · 37 features · Tokens and phrases common in academic, technical, and code-heavy documents (m…#390 · rejected 89% · 66 features · Tokens common in scientific, technical, and academic papers — citations, figu…# 11 · rejected 66% · 200 features · Formal scientific and technical prose describing methods, materials, measurem…#374 · rejected 65% · 72 features · Typical assistant responses to user questions: factual answers, how‑tos, clar…#519 · balanced · 52 features · References to lists and list-handling in technical and programming contexts (…#122 · balanced · 129 features · Common tokens and snippets from programming tutorials and code examples acros…#579 · balanced · 44 features · Generic English connective and attention-related words and phrases used in co…#305 · chosen 59% · 83 features · Table-like formatting and numeric entries in scientific, technical, and code …#696 · mixed 53/47 · 27 features · Tokens and phrases used when the assistant begins or is asked to complete its…#207 · rejected 74% · 103 features · Very common short function words, punctuation, and small code/markup tokens t…#377 · rejected 72% · 65 features · Formal patent and technical-document language describing prior art, problems,…# 98 · chosen 67% · 129 features · Calling out or citing figures, tables, and other labeled visuals in academic,…#289 · balanced · 87 features · Technical and engineering descriptions of fluid, thermal, and mechanical syst…#356 · chosen 55% · 75 features · Common small grammatical words, connective phrases, and boilerplate wording u…#145 · chosen 55% · 120 features · Common short function words, hedges, and simple transitional phrases used in …#368 · chosen 56% · 61 features · Short math question tokens and wording found in elementary arithmetic problem…#740 · rejected 73% · 19 features · Very common short function words and small grammatical tokens that appear in …#586 · chosen 59% · 43 features · Common short function words and phrasing used to indicate timing, speed, or g…#341 · chosen 54% · 77 features · Words and short phrases that signal limitation, qualification, or small disco…#499 · chosen 65% · 54 features · LaTeX and mathematical notation found in academic papers (math-mode tokens, o…#544 · rejected 73% · 49 features · Graphic sexual content and narratives—often explicit, taboo, or exploitative …#149 · mixed 49/51 · 119 features · Very common short function words, small grammatical tokens, and everyday conn…#255 · balanced · 92 features · Code snippets and programming help showing common source-code tokens, method/…#166 · rejected 68% · 102 features · Very common short function words and noisy scraped web boilerplate language#559 · mixed 51/49 · 48 features · Small code tokens and punctuation found inside source code and technical mark…#645 · rejected 66% · 36 features · Phrases and short function-word expressions used to introduce, explain, contr…#582 · rejected 71% · 44 features · Everyday user questions and requests asking for explanations, recommendations…# 83 · mixed 50/50 · 145 features · Very common short function words and everyday filler language found in inform…#772 · rejected 56% · 6 features · Transforming and simplifying algebraic expressions that combine repeated powe…# 8 · balanced · 210 features · Formal academic mathematics writing — LaTeX-style research prose with definit…#162 · balanced · 117 features · Formal expository and technical writing (encyclopedic, legal, academic, and h…#774 · chosen 63% · 11 features · Neutral, informative assistant explanations and generic content words across …#494 · balanced · 51 features · Questions and calculations converting or computing quantities, especially tim…#789 · chosen 56% · 7 features · Very common English function words and generic assistant meta-phrases used ac…#555 · rejected 56% · 37 features · Tokens and phrases common in algebraic arithmetic problems — numeric answers,…#150 · balanced · 118 features · Short code fragments and common programming patterns across multiple language…#326 · rejected 69% · 80 features · Common code patterns for accessing, assigning, updating, and iterating over p…#502 · chosen 56% · 54 features · Very common short function words and filler phrases used in everyday, instruc…#698 · chosen 60% · 27 features · Phrases and wording that describe or evaluate ease, simplicity, readiness, or…#247 · mixed 52/48 · 95 features · Words and phrases expressing size, scale, magnitude, or degree (from very sma…#154 · mixed 52/48 · 118 features · Very common short words, small function words, and everyday action/transition…#282 · rejected 68% · 87 features · Fragments and tokens from programming source code: identifiers, keywords, pun…#210 · chosen 63% · 103 features · Formatting and mention of academic citations, references, and bibliography en…#160 · balanced · 117 features · Scientific and technical text reporting numerical quantities, ranges, sample …#621 · rejected 66% · 38 features · Uses of anonymized NAME placeholders for characters, roles, and entities in p…#370 · rejected 69% · 73 features · Very common everyday words, small grammatical tokens, and routine UI/code/tec…#109 · chosen 55% · 134 features · Generic language patterns and common function words used in everyday explanat…#785 · rejected 66% · 8 features · Clear, simple explanatory text—definitions and factual explanations across sc…#314 · balanced · 83 features · Formal academic and technical prose (scientific papers, patents, and technica…#572 · balanced · 46 features · Programming problems and solutions focused on algorithms, data structures, an…#442 · balanced · 62 features · Practical programming how-to content: explanatory steps, examples, code snipp…#413 · balanced · 66 features · Instructions, examples, and structured headings used in technical or how-to d…# 0 · balanced · 555 features · Very common short words, punctuation, and generic connective tokens found acr…#576 · rejected 65% · 45 features · User prompts and assistant responses for producing written content or code—re…#128 · balanced · 126 features · News and opinion writing about political controversy, public backlash, misinf…#548 · chosen 64% · 48 features · Numeric tokens and formatted numbers as they appear in tables, scientific res…#679 · chosen 57% · 30 features · Phrases used in conversational requests, questions, and attentive replies—eve…#283 · chosen 58% · 88 features · Phrases and prompts for asking, offering, or producing written content and he…# 73 · balanced · 147 features · Formal academic mathematical writing — theorem/definition statements, proofs,…#322 · chosen 55% · 81 features · Explanations and how-to questions: instructional prompts, technical how-tos, …#780 · rejected 81% · 9 features · Markers and tokens that indicate structured, technical, or tabular document l…#398 · chosen 68% · 69 features · Numeric and statistical reporting in scientific and clinical research (tables…#105 · chosen 69% · 136 features · German-language text and common German function words, phrases, and assistant…#663 · chosen 57% · 32 features · References to containers, small personal items, and putting/storing objects (…#208 · rejected 56% · 102 features · Formal academic and scientific paper text (research articles, abstracts, figu…#404 · balanced · 67 features · Common words, phrases, and tokens used in formal academic math/physics writin…#178 · balanced · 109 features · Common short function words and formatting tokens found in technical, academi…#127 · balanced · 127 features · Common function words and generic connective tokens used in explanatory, inst…#372 · chosen 65% · 69 features · Technical discussions and code examples about low-level systems programming (…#265 · chosen 71% · 92 features · Common Russian-language assistant and informational prose: neutral conversati…#225 · rejected 56% · 100 features · Formal academic, scientific, and technical writing (research papers, patents,…#220 · balanced · 74 features · Questions and exercises that compare, rank, or sort numbers (find smallest/la…#266 · rejected 64% · 92 features · Generic words and phrases common in technical, instructional, or formal prose…#103 · balanced · 138 features · Small, common code tokens and syntax patterns found in programming examples a…#459 · balanced · 59 features · Numeric values and formatted numbers as they appear in scientific/medical tab…#252 · balanced · 95 features · Very common short function words and formatting tokens that appear across man…#379 · rejected 61% · 70 features · Short common words, UI labels, and section/headings found in web pages and en…#482 · chosen 65% · 56 features · Formal academic and scientific paper prose: LaTeX-style research writing with…#639 · chosen 57% · 36 features · Documentation and discussion around software versioning, release notes, and s…#299 · chosen 62% · 85 features · Practical programming help: tools, languages, libraries, and how-to guidance …#179 · balanced · 112 features · Language used in reviews, summaries, and creative-writing guidance—everyday w…# 7 · balanced · 215 features · Formal technical and patent-style descriptions of mechanical/electrical devic…#187 · rejected 67% · 110 features · Common high-frequency words and general-topic nouns used in encyclopedic, tec…#749 · balanced · 16 features · Very common words and names used in ordinary prose (everyday function words, …#773 · chosen 56% · 11 features · Very common everyday English words and short phrases used in general prose (f…#181 · rejected 67% · 109 features · Instruction prompts for short, structured responses and evaluations (multiple…# 3 · balanced · 251 features · General-purpose English words and phrases used in factual, technical, and bus…#118 · chosen 61% · 129 features · Mentions of URLs, web links, and web-related address fragments in technical a…#391 · balanced · 70 features · Programming and technical Q&A content with code snippets, configuration files…#280 · rejected 66% · 87 features · Tokens and phrasing related to personal names, name placeholders, and common …#486 · chosen 60% · 53 features · Language used in legal, business, and real‑estate agreements and negotiations…#395 · rejected 64% · 69 features · Formal academic and technical writing about study aims, methods, validation, …#629 · balanced · 37 features · Formal academic and technical writing in mathematics and related STEM fields …#564 · chosen 54% · 47 features · Common short function words and small phrases about movement/location or the …#195 · balanced · 107 features · Scholarly biomedical and health-science writing about physiological mechanism…#795 · balanced · 5 features · Words and phrases that add or introduce additional points in formal/informati…#267 · mixed 51/49 · 92 features · Very common short function words and small connective tokens appearing across…#256 · rejected 65% · 94 features · Formal expository and instructional writing using common function words and p…#642 · chosen 61% · 36 features · Generic explanatory and encyclopedic prose — common function words and concis…#172 · chosen 56% · 114 features · General everyday narrative language and small common words used in informal n…#568 · balanced · 33 features · Short algebra and arithmetic word-problems: solving simple linear equations, …#458 · rejected 60% · 59 features · Polite, hedging, and qualifying language the assistant uses to avoid absolute…#611 · chosen 56% · 39 features · Words and phrases about places, distances, and how near or far things are (lo…# 80 · mixed 53/47 · 140 features · Language and phrasing typical of judicial opinions and appellate court reason…#466 · chosen 55% · 58 features · Very common short English words and little connecting phrases used across man…# 42 · chosen 62% · 171 features · Text about educational institutions, academic credentials, and school-related…#651 · chosen 55% · 24 features · Questions and answers about converting or computing numbers in non-decimal ba…#714 · mixed 51/49 · 23 features · Phrases that introduce or highlight the central, main, or key part, principle…#116 · chosen 63% · 132 features · Assistant-generated Slavic-language replies and informal conversational text …#330 · chosen 68% · 79 features · Mentions of people's names and name tokens in formal, encyclopedic, academic,…#151 · balanced · 119 features · Very common short English words and connectors used across many topics (high-…#574 · chosen 57% · 46 features · Academic biomedical research and grant-writing language describing study aims…#689 · chosen 59% · 28 features · Text about reading, recommending, referencing, or linking to books/articles a…#355 · chosen 54% · 75 features · Common short function words and connective phrases used to start sentences, t…#623 · balanced · 38 features · Small punctuation, placeholders, and markup tokens commonly found inside code…#329 · chosen 65% · 79 features · Technical discussions mentioning operating systems, software versions, and de…#660 · mixed 48/52 · 33 features · Table and tabular numeric entries and headers in scientific/medical papers (d…#415 · rejected 55% · 66 features · Casual online comments and forum-style conversational fragments (posts, repli…#301 · chosen 60% · 85 features · Polite social phrasing: thanks, greetings, and courteous sign-offs used in em…#542 · rejected 63% · 35 features · Short algebra and arithmetic word problems focusing on simple polynomial/func…#627 · rejected 69% · 30 features · Tokens and phrases commonly found in Wikipedia article category lines and met…#378 · chosen 59% · 71 features · Words and phrases used to describe geographic locations, directions, and spat…#147 · mixed 49/51 · 119 features · Text about delays, omissions, missing content, and requests to stop or ignore…#671 · chosen 56% · 13 features · Boilerplate open-source license header text found at the top of source code f…#352 · rejected 58% · 70 features · Small numeric and punctuation tokens as they appear in technical, code, and t…#744 · mixed 49/51 · 19 features · Very common short function words and simple discourse connectors (basic gramm…#783 · rejected 62% · 8 features · Very common short function words and grammatical tokens used in academic, nar…#796 · chosen 59% · 5 features · Short numeric tokens, simple arithmetic expressions, and tiny common punctuat…#194 · chosen 60% · 107 features · Sexual and erotic scenes—often explicit, taboo, or non-consensual content occ…#478 · rejected 61% · 50 features · Instructions and tasks about rounding or computing numerical results to speci…#170 · mixed 47/53 · 114 features · Formal expository and instructional prose using common connective words and s…#234 · mixed 51/49 · 98 features · Very common short function words and simple transition/time phrases found acr…#311 · balanced · 81 features · Programming and technical code snippets with numeric identifiers, punctuation…#754 · mixed 49/51 · 16 features · Very common short function words and generic filler tokens used across many g…# 15 · chosen 56% · 43 features · Copyright and open-source license header boilerplate found at the top of sour…#319 · rejected 59% · 82 features · Phrases and wording typical of academic research articles (introducing backgr…#168 · chosen 60% · 115 features · Text about public events, meetings, conferences, festivals, and their schedul…# 92 · rejected 62% · 127 features · Auto-generated software documentation and its HTML navigation/structure (Java…#537 · mixed 51/49 · 50 features · Text about working, collaboration, and whether things or people function or '…#140 · chosen 61% · 122 features · Automotive content: cars, manufacturers, models, specifications, and related …#589 · balanced · 43 features · Common programming control-flow patterns: condition checks, guard/early retur…#643 · rejected 56% · 36 features · Common short function words and everyday connective language used in ordinary…# 68 · chosen 74% · 150 features · Common short function words, connectors, and everyday filler phrasing used in…#790 · chosen 59% · 7 features · Language and prompts about steps, step-by-step procedures, and incremental pr…#291 · balanced · 86 features · Tokens and short snippets from programming/configuration files — identifiers,…# 27 · chosen 62% · 186 features · Texts and discussions about Judaism and Christianity: scripture, doctrine, wo…#765 · mixed 46/54 · 12 features · Tokens from SVG/vector icon path data and other vector-graphics markup (coord…#321 · chosen 54% · 81 features · Language about future events, expectations, and conditional actions (plans, p…#490 · mixed 51/49 · 55 features · Common small connective and function words found in technical, legal, and ins…#333 · mixed 49/51 · 79 features · Common uses of the short function word “at” and related small temporal/spatia…#491 · rejected 66% · 55 features · Code patterns for conditional checks and guarding against missing/invalid val…# 79 · balanced · 146 features · Formal academic and biomedical research writing: methods, results, quantitati…#219 · balanced · 92 features · Short, common tokens and punctuation that appear in code, technical writing, …#811 · chosen 63% · 4 features · Text rich in numbers, mathematical expressions, and technical numeric notatio…#577 · balanced · 44 features · Formal scientific and technical writing patterns for reporting measurements, …#638 · chosen 55% · 35 features · References to open-source/free software, licensing, and related project or fi…# 75 · chosen 60% · 147 features · Journalistic/news article language and formatting (headlines, bylines, dateli…#271 · rejected 63% · 63 features · Short math problem prompts asking for specific numeric results (remainders, p…# 81 · chosen 61% · 145 features · Text about clothing, fashion items, and sewing/craft projects related to garm…# 63 · balanced · 153 features · Common small function words and generic connective phrases used in formal, ac…#277 · rejected 62% · 89 features · Phrases and words used to introduce, structure, and state aims or transitions…#580 · mixed 54/46 · 44 features · Very high-frequency common English function words and small grammatical token…#650 · balanced · 34 features · Formal scientific, mathematical, and technical writing with equations, numeri…#397 · chosen 64% · 69 features · Multilingual assistant conversation tokens and polite/helpful phrasing (commo…# 94 · rejected 58% · 141 features · Tropes and elements common in genre fiction summaries and plot descriptions (…#523 · balanced · 52 features · Snippets of programming source code showing function/class/method declaration…#733 · rejected 60% · 20 features · Common discourse connectives and short function words used to link, transitio…# 35 · balanced · 179 features · Very common short function words and generic connective tokens found across m…#682 · chosen 62% · 29 features · Text about immigration, visas, refugees, and legal/policy issues around migra…#441 · balanced · 62 features · Very common short function words and basic connectors (high-frequency grammat…#552 · chosen 57% · 48 features · Content about teaching, learning, and training—courses, tips, programs, and h…#302 · chosen 58% · 85 features · Biographical and news text describing people's names, roles, leadership posit…#111 · mixed 52/48 · 133 features · Words and phrases that mark time, durations, and temporal references in prose…#212 · mixed 53/47 · 91 features · Structural words and phrases found in Wikipedia articles and disambiguation p…#362 · mixed 53/47 · 74 features · References to specific dates, years, and time spans in encyclopedic, legal, m…#406 · mixed 46/54 · 67 features · Very common small words and short connective phrases used in general, technic…#419 · mixed 47/53 · 65 features · Headlines, titles, and section/heading language common to news, academic, and…#497 · chosen 61% · 54 features · Helpful, safety-conscious assistant responses giving practical advice, referr…# 22 · rejected 65% · 188 features · Common everyday words, small grammatical tokens, and routine phrases found in…#132 · balanced · 124 features · Phrases and wording used to introduce examples, links, step‑by‑step instructi…#756 · balanced · 15 features · Small tokens and punctuation in programming and technical code contexts (iden…#487 · rejected 55% · 55 features · Formal scientific and medical article language and content (clinical studies,…#286 · balanced · 88 features · Bits of programming and markup examples — common code snippets, loop/conditio…# 88 · chosen 62% · 143 features · Text about military and naval history, units, ships, deployments, and combat …#251 · rejected 59% · 95 features · Graphic sexual content and erotic scenes, often extreme/taboo (non-consensual…#351 · balanced · 76 features · Explanations and guidance about common programming concepts, patterns, and pi…#666 · mixed 48/52 · 32 features · Words and short phrases used in neutral explanatory or comparative prose (com…#517 · balanced · 53 features · Very common short function words and grammatical connectors that appear acros…# 90 · chosen 61% · 143 features · News-style reporting and analysis of international conflicts, crises, and gov…#716 · mixed 50/50 · 21 features · Combinatorial probability problems about drawing letters (probabilities of sp…# 52 · chosen 55% · 40 features · Standard open-source license header text and warranty/liability disclaimers f…# 85 · chosen 61% · 144 features · Scientific and academic astronomy/astrophysics writing and technical descript…#563 · rejected 59% · 47 features · Phrases and wording that single out, specify, or qualify particular details, …#200 · balanced · 105 features · Very common English function words and generic connective tokens that appear …#632 · rejected 59% · 37 features · Phrases and words common in assistant replies and user instruction-following …#501 · mixed 46/54 · 54 features · Language and phrases used to present, compare, and make choices or decisions …# 56 · chosen 58% · 148 features · Phrases and reasoning courts use to describe evidentiary standards, burdens, …# 99 · mixed 49/51 · 139 features · Formal expository and academic prose—general-purpose words and phrases used t…#119 · mixed 52/48 · 131 features · News-style reporting language and boilerplate (quotations, attributions, name…#318 · chosen 55% · 82 features · Short tokens and common prefixes (small function words and frequent letter cl…#670 · chosen 57% · 25 features · Text and citation formatting from U.S. judicial opinions and appellate court …#294 · balanced · 86 features · Scientific and clinical research writing reporting measurements, statistics, …#752 · rejected 62% · 16 features · Mentions of filenames, file paths, and file-related tokens in programming and…#284 · rejected 62% · 87 features · Web front-end coding: HTML/JavaScript (DOM manipulation, jQuery, UI widgets, …#755 · balanced · 15 features · Very high-frequency small English words and grammatical tokens used in genera…#222 · chosen 55% · 100 features · Words and name fragments that begin with the letters S‑ followed by a vowel c…#353 · balanced · 76 features · Numerical data and numeric table entries in scientific, medical, technical, a…#126 · chosen 55% · 128 features · Encyclopedic writing about historical places, people, institutions, and event…# 97 · chosen 58% · 139 features · Words, names, and short letter sequences starting with the letters G–Gl/Gra/G…#245 · chosen 59% · 94 features · Technical discussion and documentation about creating, processing, and displa…#573 · rejected 62% · 46 features · Common short function words, punctuation, name fragments, and small formattin…#430 · rejected 62% · 64 features · Greetings, salutations, and brief polite openings used to start conversations…#221 · chosen 57% · 100 features · Phrases where the assistant introduces itself, states limits or safety refusa…#522 · mixed 50/50 · 50 features · Common small function words and short connective tokens (especially uses of “…#493 · mixed 50/50 · 54 features · Tokens and punctuation patterns common in technical, academic, and patent tex…# 47 · chosen 61% · 169 features · Mentions of geographic places, location words, and travel/encyclopedic place …#135 · chosen 59% · 123 features · Product descriptions and technical details for consumer mobile and home elect…#520 · rejected 62% · 52 features · Academic biomedical and life-science prose reporting experimental methods, mo…#631 · balanced · 27 features · Short math and arithmetic problem prompts and small-function words common in …#641 · mixed 51/49 · 36 features · Phrases that use “let” to invite, prompt, transition, or soften statements (c…#293 · mixed 49/51 · 85 features · Common connective and formal phrasing used to state reasons, bases, and proce…#509 · mixed 48/52 · 53 features · Words and phrases that describe permanence, lasting change, or habitual/ongoi…#515 · rejected 64% · 50 features · Short math problem language: elementary arithmetic, integer operations, compa…#781 · mixed 52/48 · 9 features · Everyday technical help and troubleshooting language—practical programming fi…#538 · rejected 58% · 50 features · Phrases expressing apology, regret, or polite refusal (including empathetic a…#108 · chosen 59% · 133 features · Descriptions and discussions of weather, temperature, seasons, and plant grow…#227 · mixed 48/52 · 100 features · Very common small words, connectors, and routine phrases used in informationa…#424 · rejected 65% · 65 features · Common small function words and everyday connective language found across new…# 96 · balanced · 140 features · Scientific, clinical, and technical reporting language that expresses measure…#313 · chosen 60% · 83 features · Text about places, countries, regions, and geographic context (travel, demogr…#731 · rejected 60% · 20 features · Statements comparing options that highlight benefits, pros, trade‑offs, and w…#236 · rejected 57% · 96 features · Formatting and numbering tokens common in academic, technical, or code-like d…#231 · balanced · 99 features · Tokens and syntax from web and configuration markup — HTML/JSX tags, JSON/XML…#513 · mixed 45/55 · 52 features · Boolean/state fields and short structural tokens common in code, configs, and…#776 · mixed 52/48 · 11 features · Formal academic and technical writing (scientific, medical, and professional …#238 · rejected 64% · 97 features · Tokens and phrases commonly found across diverse English texts—generic functi…#508 · mixed 49/51 · 53 features · Phrases about remembering, memories, and reminders (nostalgia, 'remember', no…#444 · balanced · 59 features · Source-code files and developer-facing documentation (comments, docblocks, li…#152 · balanced · 119 features · Very common short function words and small grammatical tokens used across man…#715 · chosen 56% · 23 features · Short first‑person and second‑person conversational phrases and common functi…#809 · chosen 57% · 4 features · Everyday conversational words and short phrases used in casual narratives (pe…# 50 · balanced · 168 features · LaTeX-formatted mathematical and scientific expressions common in research pa…#719 · mixed 52/48 · 22 features · Requests for concise summaries, overviews, or brief explanatory how‑to descri…#767 · mixed 47/53 · 12 features · Very frequent short English words and punctuation used as generic connectives…#394 · chosen 54% · 66 features · Standard ethical, funding, and boilerplate statements found in academic and b…#658 · chosen 55% · 30 features · Oligonucleotide sequences and PCR/qPCR primer details in molecular biology me…#722 · chosen 54% · 21 features · References to U.S. states, cities, local places, and regional/local news or l…#112 · rejected 57% · 133 features · Methods, sampling, and reporting language from medical and public-health rese…#695 · chosen 57% · 27 features · Encyclopedic and administrative content about East Asian places, people, lang…#146 · mixed 52/48 · 120 features · Phrases and expressions that refer to time, duration, or temporal points (whe…#613 · chosen 60% · 39 features · Encyclopedic and news-style content about places, people, and institutions ac…#405 · balanced · 50 features · Combinatorics word problems asking for probabilities of drawing specific lett…# 95 · chosen 55% · 140 features · Technical networking and telecommunications topics: protocols, configuration,…#676 · chosen 56% · 30 features · Marketing and local business promotional copy highlighting services, expertis…#617 · mixed 53/47 · 24 features · Language and citation conventions found in U.S. appellate and judicial opinio…# 91 · chosen 58% · 142 features · Common tokens and short word fragments (place names, proper nouns, short func…#505 · rejected 55% · 54 features · Common short function words and negation tokens used to qualify, limit, or de…# 87 · mixed 51/49 · 143 features · Text about buying, selling, shipping, and practical commercial transactions (…#120 · rejected 64% · 129 features · Common short function words and connective phrases used across news, academic…#485 · rejected 58% · 55 features · Comparative and contrast language used in technical, scientific, and formal w…#616 · rejected 57% · 39 features · Tokens and punctuation that mark structure within code, JSON, XML, and notebo…# 9 · chosen 59% · 206 features · Legal and contractual texts about liability, corporate/government entities, a…#495 · chosen 56% · 54 features · Talk about time—how people spend, have, or take short periods (sleep, naps, w…#531 · balanced · 51 features · Formal informational and technical writing (academic, scientific, patents, an…# 2 · chosen 56% · 247 features · Texts about plants, crops, and wild animals—botanical names, cultivation, pes…#242 · rejected 59% · 96 features · User prompts and roleplays trying to override assistant rules—jailbreaks, unc…# 38 · chosen 62% · 176 features · Explicit sexual content, often violent or non-consensual, focusing on sexual …#276 · rejected 55% · 88 features · Common high-frequency function words and generic connective phrasing that app…#157 · chosen 57% · 107 features · Low-level system and device driver source code and documentation (kernel/firm…#274 · chosen 59% · 90 features · Scholarly neuroscience texts about brain systems, sensory/motor function, mem…#360 · chosen 56% · 74 features · Discussions and code about colors, color properties, and background styling i…#106 · rejected 63% · 136 features · Common short words, pronouns, punctuation, and everyday discourse markers use…#385 · balanced · 71 features · Structured lists, outlines, and step-by-step instructional content (bulleted/…#335 · chosen 57% · 78 features · Generic everyday language and common instructional/commercial phrasing (routi…#174 · mixed 50/50 · 113 features · Everyday expressions of personal reaction, interest, and polite announcements…#602 · mixed 50/50 · 40 features · Concepts and mentions of 'auto' / self-directed or automated systems and 'sel…#793 · chosen 58% · 6 features · Very common short words and place/metadata tokens that appear in general news…#807 · mixed 50/50 · 4 features · Short common words and punctuation used as generic connectors in running text…#461 · chosen 54% · 59 features · Polite assistant responses about its abilities, limits, and offering help (fr…#298 · rejected 71% · 85 features · Generic high-frequency small words and punctuation used as connective 'glue' …#209 · chosen 59% · 101 features · Polite assistant language offering help, follow-ups, and brief self-descripti…# 54 · chosen 60% · 159 features · Practical command-line, scripting, packaging, and system administration tasks…#411 · chosen 55% · 66 features · Topics about healthcare delivery, hospitals, and health policy (care settings…#503 · mixed 51/49 · 54 features · Polite, helpful assistant phrasing inviting clarification, follow‑up, or offe…# 19 · chosen 54% · 190 features · News and reference text about elections, political officeholders, campaigns, …#232 · rejected 75% · 98 features · Short tokens, symbols, and code-like fragments common in technical, scientifi…#614 · mixed 47/53 · 35 features · Common small grammatical words and short instruction-related phrasing typical…#655 · balanced · 34 features · Requests and outputs for basic integer arithmetic tasks—especially finding or…#504 · mixed 49/51 · 54 features · Common short function words and casual conversational phrases centered on 'ge…#747 · chosen 56% · 17 features · Phrases and usages that begin with or revolve around the word “under,” signal…#775 · chosen 55% · 11 features · Instructions and questions asking to round or format numbers to a specified p…#433 · chosen 55% · 62 features · User questions and brief conversational prompts asking for help, opinions, cl…# 78 · rejected 61% · 146 features · Common sentence-level connective and transitional words/phrases used to link,…#196 · balanced · 107 features · Laboratory and industrial chemical synthesis procedures and purification/work…# 72 · mixed 46/54 · 147 features · Casual multilingual social-media style text (emoji-rich, noisy or non-Latin s…#600 · rejected 63% · 41 features · Code patterns for attaching and handling event callbacks/listeners in user-in…#358 · balanced · 74 features · Technical and academic writing tokens and formatting common in research, code…# 32 · balanced · 181 features · Reporting and describing statistical methods, results, and significance (p‑va…#125 · chosen 58% · 128 features · News and legal writing about criminal charges, arrests, prosecutions, sentenc…#610 · balanced · 36 features · Texts about ordering and sorting things—instructions, questions, or code deal…# 24 · chosen 59% · 188 features · Medical and surgical descriptions of human anatomy, injuries, vascular and mu…#243 · balanced · 70 features · Middle-school to high-school style algebra and arithmetic problems: simplifyi…#753 · chosen 60% · 16 features · Markers, names, and links used in online discussion replies and quoted forum/…#139 · chosen 55% · 122 features · Phrases and responses enforcing safety rules—refusing or restricting harmful,…#561 · balanced · 41 features · Questions and problems about divisibility, remainders, and numeric comparison…#328 · chosen 57% · 79 features · Discussions of gender, sexual orientation, and experiences of marginalized ra…#421 · mixed 48/52 · 65 features · Phrases and verbs centered on 'take'—actions of taking, undertaking, acceptin…#553 · mixed 50/50 · 47 features · Common short function words and simple temporal/transition phrases used to sh…# 53 · mixed 53/47 · 164 features · Descriptions and listings of buildings and homes, including architectural fea…#481 · rejected 57% · 56 features · Short conversational markers and tasks for confirming factual consistency (br…#806 · rejected 58% · 4 features · Finding or selecting numeric coefficients when simplifying or rearranging alg…#393 · mixed 48/52 · 69 features · Words and phrases about leaving, staying, and keeping or storing things (rema…#477 · rejected 73% · 47 features · Short, common English words and function tokens that glue sentences together …#587 · rejected 56% · 42 features · Short function words, punctuation, and small numeric tokens used in technical…#465 · mixed 53/47 · 58 features · Language used in clinical research papers to report patient outcomes, surviva…#235 · rejected 65% · 98 features · Very common short function words and generic boilerplate phrasing found acros…#778 · chosen 54% · 10 features · Formal academic and scientific prose — common small words, connectors, and re…#416 · mixed 50/50 · 66 features · Common short function words and everyday discourse connectors used to link, q…#134 · chosen 60% · 124 features · Text and instructions for graphical user interfaces—buttons, menus, dialogs, …#258 · rejected 58% · 94 features · Common small function words and everyday connective phrases found throughout …#101 · chosen 57% · 138 features · Content about online promotion, social media, websites, and digital marketing…#349 · balanced · 76 features · Practical step‑by‑step programming and developer how‑to instructions (code ex…#450 · mixed 53/47 · 60 features · Phrases and wording writers use to comment on how things seem, look, or appea…#366 · mixed 49/51 · 73 features · Common small function words and simple discourse connectors used to start, co…#153 · rejected 64% · 119 features · Common everyday English words and short phrases used to connect, refer to peo…#628 · mixed 48/52 · 37 features · Words and terms that begin with short common letter sequences or prefixes (e.…#136 · chosen 57% · 123 features · References to dates, years, and time periods in articles, biographies, legal …#547 · chosen 55% · 49 features · Descriptions, guidance, and assessments of physical movement, posture, and ex…#159 · rejected 59% · 108 features · Language and structure of appellate and trial court opinions — procedural his…#263 · mixed 50/50 · 92 features · Phrases and passages expressing uncertainty, lack of knowledge, or not knowin…#453 · mixed 50/50 · 60 features · Common small function words and generic connective phrasing used in everyday …#596 · mixed 53/47 · 42 features · Technical and scientific writing (academic papers, methods, and discussion of…#226 · rejected 67% · 100 features · Generic high-frequency English words and small grammatical tokens used across…#438 · mixed 49/51 · 63 features · Sexualized descriptions of people and erotic roleplay, often explicit and som…#677 · chosen 54% · 30 features · Polite well-wishes and closing phrases used by writers or assistants (e.g., '…#724 · chosen 54% · 21 features · Words and phrases expressing continuation, ongoingness, or that something rem…#569 · rejected 60% · 46 features · Texts and phrases about homes, living situations, and the concept of life/bei…#704 · rejected 63% · 17 features · Algebra and calculus problems focused on simplifying power/exponent expressio…#269 · chosen 61% · 87 features · Code snippets and technical discussion about statically typed programming lan…#524 · chosen 58% · 52 features · Legal and constitutional rights and doctrines discussed in court opinions and…#451 · mixed 49/51 · 60 features · Positive, reassuring, or evaluative phrasing used in everyday advice, reviews…#787 · rejected 62% · 8 features · Technical code and configuration text: program snippets, logs, numeric parame…#536 · mixed 48/52 · 50 features · Formal academic and professional writing (scientific, legal, and technical pr…#791 · mixed 52/48 · 7 features · Generic short common words and promotional/descriptive phrases found in every…#701 · mixed 48/52 · 27 features · Phrases about trying, attempting, or encouraging people to try (efforts, expe…#717 · rejected 58% · 23 features · Methods and materials descriptions in scientific/biomedical papers (experimen…#730 · mixed 51/49 · 20 features · Phrases and transitions that refer to time—past events, recent findings, or f…#606 · mixed 48/52 · 39 features · Words and phrases related to passing, transmitting, sharing, or handing somet…#418 · mixed 48/52 · 65 features · Generic common words, names, and short topical tokens found in diverse encycl…#454 · rejected 61% · 59 features · Clinical and biomedical research writing describing patients, study participa…#163 · chosen 57% · 116 features · Words and names starting with the letters T–Tr (short common tokens, place/pe…#130 · chosen 57% · 124 features · Methods and equipment details for laboratory experiments and biomedical imagi…#675 · mixed 50/50 · 30 features · Text about dividing, splitting, or grouping things into parts, categories, se…# 93 · mixed 53/47 · 142 features · Short words, prefixes, and name fragments that commonly appear in varied prop…#342 · mixed 52/48 · 76 features · Tokens and short word fragments commonly found in encyclopedic, formal, and t…# 21 · rejected 64% · 189 features · Everyday connective language and hedging used in narratives, explanations, an…#344 · mixed 54/46 · 76 features · Text about visual art, artists, exhibitions, museum collections, and image/co…#254 · balanced · 94 features · Very common small function words and short grammatical connectors that appear…#143 · chosen 56% · 121 features · Academic and public-health research writing about mental/physical health, epi…#634 · mixed 53/47 · 36 features · Academic and news reporting of population-level statistics, prevalence, rates…#726 · chosen 54% · 21 features · Textual mentions of physical sizes, dimensions, and numeric measurements (uni…#644 · rejected 54% · 36 features · Uses of the word “from” and other common small function words in headings, at…# 45 · rejected 63% · 169 features · Explicit sexual content, often violent or taboo (non-consensual, incest, mino…#674 · mixed 49/51 · 30 features · Words and text snippets featuring 'comp'/'com' beginnings—terms like company,…#521 · chosen 56% · 52 features · Text about peoples, cultures, and their histories—including indigenous groups…#386 · chosen 59% · 70 features · Mentions of countries, places, and local names in news, encyclopedia, or regi…# 20 · mixed 52/48 · 179 features · Legal court opinions and filings focusing on procedural history, motions, dea…#532 · rejected 58% · 45 features · Place names and postal/address fragments found in formal documents (addresses…#369 · balanced · 68 features · Technical and programming text: code snippets, math/formal definitions, and t…# 28 · chosen 58% · 185 features · Content about movies, TV shows, actors, production credits, and related popul…#392 · balanced · 70 features · Concise, practical Stack Overflow–style programming help: code fixes, how-tos…#758 · rejected 57% · 10 features · Requests and operations involving symbolic derivatives and simple polynomial …# 66 · rejected 56% · 151 features · Casual everyday language and personal-expression tokens used in informal conv…#387 · mixed 50/50 · 70 features · General-purpose tokens and headings found in technical, encyclopedic, and web…#435 · rejected 59% · 60 features · Language for comparing, ranking, or calling something the most/least (superla…#476 · mixed 52/48 · 42 features · Objective-C, Swift, and iOS/macOS developer code and API patterns (UIKit/AppK…#525 · mixed 47/53 · 52 features · Academic and clinical biomedical writing about diseases, pathology, treatment…#246 · mixed 47/53 · 95 features · Academic, scientific, or technical writing and terminology (formal research-s…#365 · chosen 60% · 73 features · Encyclopedic historical writing—overviews and discussions of past periods, ev…#489 · balanced · 54 features · Very common small function and connective words used to link ideas, indicate …#527 · balanced · 51 features · Code snippets showing conditional checks and guard clauses (if/elif/else, mem…# 89 · rejected 58% · 143 features · Common short function words and everyday connective phrases used to link, ref…#363 · mixed 48/52 · 73 features · Formal or technical writing in academic, scientific, medical, and technical d…#428 · balanced · 64 features · Practical programming and developer how-to content: installation, code exampl…#558 · balanced · 48 features · Very common short function words and grammatical tokens appearing in formal, …# 4 · mixed 52/48 · 230 features · News and commentary about sports games, teams, and player performance (recaps…#315 · mixed 49/51 · 83 features · Text that reports or references numerical proportions, percentages, and other…#471 · chosen 55% · 56 features · Text about light, brightness, and optical/vision topics (sunlight, illuminati…#498 · balanced · 54 features · Common small function words and tokens used in technical, numerical, and inst…# 71 · balanced · 149 features · Technical writing about electronic and optoelectronic hardware, fabrication, …#788 · mixed 50/50 · 7 features · Technical and academic texts with numbers and inline reference markers (measu…#514 · mixed 52/48 · 53 features · Scientific and medical research text about human and animal anatomy, diseases…# 59 · chosen 57% · 161 features · Technical descriptions of industrial organic chemistry: synthesis, production…#685 · balanced · 29 features · Code and developer documentation: programming examples, how‑to instructions, …#720 · mixed 51/49 · 22 features · Words and phrases about hiding, covering, or filling spaces—concealment, fill…#727 · balanced · 20 features · Common small function words and hedging/temporal phrases used in general info…#540 · rejected 56% · 47 features · Source code tokens and common programming syntax across multiple languages (i…#203 · balanced · 104 features · Programming examples and discussions about looping and indexing over arrays/c…#585 · mixed 47/53 · 43 features · Uses of the root “act/activ/active” and related action-language across legal,…#102 · mixed 46/54 · 135 features · Texts about chemicals, chemical companies, and industrial uses—descriptions o…#214 · mixed 53/47 · 102 features · Text about transportation and travel infrastructure (stations, routes, vehicl…#687 · mixed 53/47 · 28 features · Assistant safety and refusal boilerplate, plus user jailbreak prompts trying …#688 · mixed 51/49 · 28 features · Words and phrases related to play, fun, joking, and playful behavior in varie…#182 · balanced · 106 features · Technical discussion and code for 2D/3D graphics and game development (render…# 69 · chosen 56% · 148 features · Supportive self-help and spiritual guidance encouraging personal growth, resi…#148 · rejected 64% · 119 features · Common small function words and connective phrases used in questions, technic…#337 · chosen 56% · 78 features · Mentions of people’s ages, age groups, and age-related demographic information#156 · mixed 50/50 · 115 features · Developer-focused API/SDK and cloud infrastructure code snippets, configurati…#124 · mixed 53/47 · 128 features · Tokens that are parts of personal and place names or other capitalized proper…#384 · rejected 56% · 71 features · Common connective and helper phrases used to explain, advise, or describe act…#443 · rejected 60% · 53 features · C and C++ source/header file structure and boilerplate tokens (namespaces, in…#107 · rejected 56% · 135 features · Phrases where the assistant explains its identity, capabilities, limits, or l…#183 · balanced · 108 features · Tokens and small snippets from programming source code (identifiers, types, d…#745 · balanced · 15 features · Questions and answers about finding the n-th term or simplifying polynomial e…#743 · rejected 60% · 18 features · Short tasks asking the model to judge everyday behaviors as “it's okay,” “it'…# 74 · balanced · 147 features · Questions and code about relational databases and SQL (queries, schema, error…#184 · mixed 50/50 · 56 features · Boilerplate text and header language from U.S. appellate court opinions (cita…#779 · mixed 48/52 · 9 features · Academic and scientific writing (formal papers in math/physics/condensed-matt…#647 · mixed 53/47 · 35 features · Descriptions, reviews, and sensory/production discussion of alcoholic beverag…#432 · chosen 58% · 62 features · Common tokens and phrasing found in programming documentation and code exampl…#595 · mixed 51/49 · 41 features · Tokens and phrases common to personal blogs and informal dated posts (everyda…#448 · mixed 51/49 · 60 features · Everyday conversational words and pronouns used in instructions, clarificatio…#535 · mixed 48/52 · 50 features · Formal academic and clinical research prose (scientific article language, esp…# 1 · chosen 58% · 253 features · Text about sports events and results—especially horse and motor racing, match…#529 · rejected 63% · 51 features · Short, high-frequency tokens and identifiers common in programming and techni…#304 · rejected 58% · 71 features · Mobile app user interface and lifecycle development (Android and iOS layout X…#543 · mixed 49/51 · 49 features · Words and names containing the letter sequence “ol/ol-” (proper names, scient…#712 · balanced · 24 features · Scientific, technical, and code text that uses numbers and short common words…#510 · mixed 49/51 · 53 features · Common small function words and generic phrasing that appear across formal, i…#541 · rejected 61% · 35 features · Short math problems and algebraic simplifications (school-level arithmetic, p…#192 · chosen 55% · 107 features · Clinical and biomedical texts reporting blood-based laboratory measurements a…#603 · mixed 53/47 · 36 features · Programming source files and code snippets across many languages, especially …#718 · rejected 57% · 22 features · Language about achieving goals, success, effort, and ways to accomplish thing…#530 · chosen 57% · 49 features · Code-related testing and test suites: unit/integration tests, test frameworks…#697 · chosen 60% · 26 features · Snippets of programming source code and build/compile-related tokens across l…#607 · chosen 60% · 39 features · Texts about firearms, shooting, explosions, and related violent or fire/arson…#803 · mixed 51/49 · 4 features · Text about historical, political, and medical topics using common small words…#367 · chosen 55% · 73 features · Words and names starting with the letters 'L'+'o/Li/Lu/Les' appearing across …#142 · rejected 54% · 121 features · Everyday conversational and instructional language—polite offers, requests, h…#612 · rejected 61% · 39 features · Practical advice and reminders phrased as cautious, helpful instructions (ens…#240 · chosen 55% · 96 features · Problems and notations involving numbers and numeric sequences in technical, …#188 · balanced · 108 features · Formal technical and academic prose introducing or linking ideas, methods, re…#786 · mixed 47/53 · 6 features · Flutter and Dart UI code patterns and typical widget/state usage examples#382 · chosen 58% · 70 features · Formal, encyclopedic and literary prose from older printed texts—biographical…#215 · mixed 53/47 · 102 features · Mentions of government, public institutions, and formal official bodies and r…#648 · mixed 51/49 · 31 features · Short math and technical snippets: small-answer arithmetic/algebra problems, …#757 · mixed 50/50 · 12 features · Auto-generated source file headers, comments, and the surrounding code fragme…#462 · rejected 57% · 59 features · Informal spoken and quoted language in news, interviews, and media (transcrib…#257 · rejected 54% · 94 features · Common small connective words, punctuation, and short function tokens used to…#155 · mixed 50/50 · 118 features · Words, names, and technical terms beginning with the letter F across diverse …#770 · mixed 49/51 · 11 features · Very high-frequency small English words and short functional tokens (common a…#357 · rejected 58% · 65 features · Front-end web pages and templates: HTML documents with CSS, meta tags, script…#186 · balanced · 110 features · Technical academic writing in theoretical and experimental high-energy/partic…#684 · rejected 55% · 29 features · Academic and scientific methods/results wording describing study design, meas…#104 · mixed 48/52 · 136 features · Content about games and gaming: video games, tabletop/RPG mechanics, card gam…#692 · mixed 48/52 · 28 features · Local and national public-service, government, and institutional informationa…#657 · balanced · 33 features · Small code/math tokens and index placeholders appearing in technical formulas…#300 · chosen 55% · 83 features · Technical documentation and troubleshooting for software projects: file paths…# 16 · mixed 47/53 · 185 features · Promotional ‘About the Company’ introductions for chemical manufacturers and …#506 · chosen 56% · 53 features · Narrative scene descriptions focusing on locations, sudden sounds, and sensor…#439 · mixed 51/49 · 63 features · Texts about material damage, wear, and protective measures—how things break, …#597 · rejected 61% · 42 features · Short, common words and connectors used in simple factual explanations and br…#534 · rejected 58% · 51 features · Wikipedia-style encyclopedic article text (biographies, places, organizations…#565 · chosen 55% · 47 features · Text about rating, scoring, ranking, and reviews (numerical or categorical ev…#711 · mixed 50/50 · 24 features · Common short function words and connective tokens used across everyday and na…# 70 · mixed 53/47 · 149 features · Travel and tourism descriptions: places, attractions, outdoor activities, and…#618 · rejected 59% · 39 features · Very common function words, short connectors, and routine punctuation found i…#668 · chosen 55% · 32 features · Topics about cleaning, washing, hygiene, and household personal-care products…#402 · rejected 56% · 67 features · Building and using HTTP APIs: requests, headers, methods, URLs, responses, an…#686 · mixed 47/53 · 28 features · Common short function words and filler tokens that structure everyday sentenc…#380 · mixed 53/47 · 44 features · U.S. court opinions and legal case documents (headings, citations, procedural…# 76 · chosen 60% · 147 features · Mentions of dates and times in factual or informational writing (calendar dat…#759 · mixed 52/48 · 14 features · Mentions of a particular area, field, location, or 'on the ground' context (r…#325 · mixed 50/50 · 58 features · Citations and prior‑art references and boilerplate language commonly found in…# 49 · chosen 56% · 168 features · News and legal reporting about crimes, arrests, violent incidents, deaths, an…#551 · rejected 57% · 48 features · Very common small words and generic content words used in everyday prose and …#474 · chosen 54% · 38 features · LaTeX and Math typesetting in academic and technical documents (equations, pa…#581 · mixed 52/48 · 43 features · Assistant messages about language abilities, translations, and switching or o…#591 · rejected 55% · 41 features · Code examples and snippets from programming Q&A, especially printing/logging,…#436 · chosen 55% · 58 features · Mentions of towns, cities, local places, and related local government or comm…#622 · chosen 56% · 38 features · Discussion of concurrency and parallelism patterns: threads, locks, pools, qu…# 5 · mixed 54/46 · 220 features · Cooking and food-related recipe content (instructions, ingredients, preparati…# 84 · rejected 64% · 144 features · Very common small words and grammatical tokens used in neutral explanatory or…#275 · mixed 47/53 · 89 features · First-person expressions and conversational phrasing used to state beliefs, e…#702 · chosen 55% · 27 features · Words, names, and tokens containing the letters Z/zo/zu/zen (proper names, br…#324 · mixed 47/53 · 80 features · Generating poems, song lyrics, and other creative lyrical writing (including …#546 · mixed 52/48 · 49 features · Biomedical texts about cancer: tumors, pathology, metastasis, biomarkers, and…#667 · mixed 53/47 · 32 features · Scholarly, medical, or technical writing and headings—especially academic/cli…#665 · mixed 49/51 · 32 features · Words and phrases about places, locations, and asking/indicating where someth…#224 · balanced · 100 features · Text snippets containing numbers, measurements, units, and quantitative value…#237 · chosen 55% · 98 features · Encyclopedic biographical text about people's lives, careers, dates, and plac…#347 · mixed 49/51 · 76 features · Methods and experimental descriptions from biomedical research articles (anim…#460 · balanced · 35 features · Short arithmetic and algebra word problems asking to solve simple linear equa…#583 · balanced · 44 features · Phrases and tokens from clinical and biomedical research reports describing p…#171 · mixed 49/51 · 114 features · Common short substrings, name fragments, and small function words that appear…#272 · mixed 49/51 · 87 features · Proper names, company/place names, and common short function words found in f…#594 · balanced · 42 features · Very high‑frequency English function words and short grammatical tokens used …#705 · mixed 47/53 · 26 features · Statistical methods and reporting of outcomes, group comparisons, and covaria…#694 · mixed 48/52 · 27 features · Language commonly used in organizational or technical support, promotion, and…#422 · rejected 58% · 65 features · Frequent small function words and common connective phrases used across factu…#768 · chosen 62% · 9 features · Website and document placeholder text and template filler (e.g., Lorem ipsum,…#624 · mixed 48/52 · 38 features · Requests for written professional content and structured summaries (company i…# 29 · mixed 52/48 · 184 features · Encyclopedic entries and metadata about music (artists, albums, performances,…#625 · rejected 55% · 38 features · Phrases and tokens common in academic and technical papers—section headings, …# 55 · mixed 47/53 · 159 features · Common short tokens and assistant-style replies across multiple languages—pol…# 64 · mixed 53/47 · 123 features · Footer category and metadata lines from Wikipedia article pages (the category…#123 · chosen 56% · 129 features · Place names and geographic references in encyclopedia-style and news texts (c…#249 · balanced · 81 features · Formal academic and scientific writing—especially biomedical, biological, and…# 36 · mixed 48/52 · 165 features · Generic website boilerplate and legal/disclaimer copy found on commercial sit…#345 · rejected 56% · 76 features · Common small function words and short connective phrases used to link, transi…# 60 · rejected 58% · 158 features · Common function words and generic informational phrasing found in encyclopedi…#100 · balanced · 106 features · Tasks involving polynomials and basic symbolic algebra — simplifying, extract…#412 · balanced · 56 features · Solving and simplifying algebra problems about polynomials, roots, factoring,…#288 · rejected 62% · 80 features · Website and blog content elements—sidebars, footers, metadata, listings, and …#408 · rejected 54% · 67 features · Phrases and small words used to instruct, urge, or describe actions and oblig…#403 · chosen 58% · 68 features · Technical help and how-to responses in Romance languages, especially programm…#633 · mixed 48/52 · 37 features · References to official paperwork, signatures, and procedural document require…#693 · mixed 48/52 · 28 features · Text about organizations, groups, and their structures, history, and activiti…#746 · chosen 56% · 17 features · Casual to aggressive swear words, insults, and explicit bodily-excretion lang…#496 · mixed 48/52 · 54 features · Text about awards, honors, recognitions, and formal distinctions given to peo…#554 · rejected 54% · 48 features · Very common short English function words and handful of generic 'one/single' …#464 · mixed 54/46 · 58 features · Discussion of laws, regulations, government policy, and compliance requirements# 6 · balanced · 217 features · News and reports about finance, markets, and government fiscal/economic policy#769 · mixed 50/50 · 12 features · Mentions of searching and search UI—advice to look things up, search boxes, r…#488 · mixed 49/51 · 48 features · Questions that compute probabilities of drawing specific letters, sequences, …#584 · chosen 56% · 44 features · Mentions of organizations, institutions, and formal roles (companies, agencie…#810 · mixed 51/49 · 4 features · Phrases expressing caution, doubt, or reserved expectations about people or o…#445 · mixed 49/51 · 61 features · Phrasing about trends, progress, adoption, and changing popularity of technol…#332 · mixed 49/51 · 71 features · Development questions and code snippets about .NET/C# application UI, project…#593 · mixed 52/48 · 42 features · References to geographical place names and proper nouns, especially those sta…#480 · mixed 48/52 · 56 features · Assistant statements noting its limits, lack of consciousness, or inability t…#296 · mixed 47/53 · 85 features · Advice, tools, and warnings for keeping people, data, and systems safe and pr…#191 · rejected 58% · 108 features · Inline citation markers and reference-bracket tokens commonly used in academi…#361 · mixed 51/49 · 74 features · Common running-text words and named entities found in encyclopedic, news, and…#636 · rejected 57% · 37 features · Very high-frequency small words, punctuation, and common function tokens that…#259 · rejected 55% · 92 features · Polite assistant introductions, self-descriptions, and safe refusals that exp…#426 · mixed 52/48 · 63 features · Everyday conversational words and short phrases used to make, contrast, or di…#590 · rejected 65% · 43 features · Very high-frequency small English words and tiny punctuation/formatting token…#201 · rejected 57% · 105 features · Text about romantic and sexual attraction, relationships, and social/sexual e…#479 · rejected 55% · 56 features · Short words, name fragments, and common substrings that appear across multili…#690 · mixed 49/51 · 28 features · Text about industrial or chemical production, manufacturing processes, and co…#169 · balanced · 92 features · Short algebra and arithmetic word problems: polynomial manipulation, derivati…#250 · rejected 61% · 95 features · Numbered or ordered steps and list markers used in instructions, procedures, …# 67 · mixed 50/50 · 111 features · Standard patent boilerplate introducing the invention’s field, objectives, an…#709 · rejected 61% · 25 features · Practical programming help and troubleshooting for developers (Stack Overflow…#804 · mixed 48/52 · 4 features · Very common sentence-level tokens and connective words that mark ordinary sen…#431 · mixed 52/48 · 46 features · Sections of scientific papers describing X‑ray crystallography experiments, c…#737 · rejected 63% · 20 features · Small syntactic tokens and identifier patterns inside programming source code…#507 · rejected 60% · 53 features · Descriptions and references to methods, approaches, paths, and step-by-step p…#500 · chosen 55% · 53 features · Methods and materials details in biomedical research papers (statistical soft…#470 · rejected 55% · 58 features · Mentions of dates, times, and scheduling in code and documentation (calendar …#193 · balanced · 69 features · Algebraic arithmetic and simplification problems—combining exponents, roots, …#708 · chosen 54% · 25 features · Text about fights, combat, and competitive conflict—especially combat sports …#264 · mixed 52/48 · 92 features · Tokens that begin with the letters H–Ha/He/Hi/Ho/Ho* — common names, places, …#440 · mixed 47/53 · 62 features · Common small words, name fragments, and frequent short tokens found across ge…#735 · mixed 52/48 · 19 features · Code and discussion about building graphical user interfaces — widgets, layou…#699 · rejected 58% · 24 features · Phrases and wording about limits, restrictions, and stated limitations (in te…#516 · mixed 51/49 · 52 features · Medical and biological topics related to pregnancy, childbirth, newborns, and…#764 · mixed 51/49 · 12 features · Writing about celebrities, fame, and cultural impact—tabloid-style coverage a…# 82 · mixed 52/48 · 128 features · Runtime and compile-time error messages, logs, and stack traces from programm…#339 · chosen 53% · 74 features · Contact and communication information in emails and websites (phone numbers, …#640 · balanced · 36 features · Very common English function words and generic discourse connectors used in n…#161 · balanced · 117 features · Programming texts about handling, searching, validating, and manipulating str…#700 · mixed 48/52 · 27 features · In-text academic citations, reference lists, and author name tokens in schola…#213 · rejected 61% · 97 features · High-entropy technical identifiers and encoded data found in software, logs, …#189 · mixed 53/47 · 108 features · Health and nutrition content about diet, nutrients, body weight, and related …#673 · rejected 54% · 31 features · Common short function words and everyday connective phrases used in neutral i…#539 · mixed 47/53 · 50 features · Marketing and explanatory content about enterprise software, technology produ…#455 · mixed 45/55 · 46 features · Short arithmetic and algebra problems and requests (basic integer/fraction ca…#635 · balanced · 37 features · Short numeric and alphanumeric tokens used as indices, IDs, counts, or small …#467 · balanced · 58 features · Code and examples for reading, writing, managing, and inspecting files and di…#376 · rejected 57% · 72 features · Common words and phrasing found in formal, technical, and academic documents …#261 · mixed 46/54 · 87 features · Programming source code tokens and common syntax patterns found in C-like and…#167 · chosen 56% · 110 features · Date and time expressions and timestamps in texts, logs, and metadata (times,…#306 · mixed 48/52 · 69 features · Website and forum UI boilerplate and community engagement prompts (registrati…#216 · mixed 48/52 · 102 features · Short, everyday words and phrases used to ask questions, make transitions, or…#158 · balanced · 118 features · Technical descriptions of materials, their synthesis, properties, and manufac…#253 · mixed 46/54 · 94 features · Language that describes people's feelings, emotional reactions, and sensation…#285 · chosen 59% · 88 features · Scientific and technical writing about chemicals, materials, and laboratory m…#164 · rejected 53% · 116 features · Scientific discussions of immune system components, immune responses, and lab…#468 · mixed 50/50 · 58 features · Technical programming help and code-related documentation (developer Q&A, cod…#373 · mixed 53/47 · 72 features · Information about chemical hazards and practical safety precautions for handl…#309 · rejected 56% · 84 features · Instructions and discussion about making changes, adjustments, or switching s…#113 · mixed 51/49 · 117 features · Formal legal and regulatory texts with statutory citations, section markings,…#338 · balanced · 78 features · Common small function words, connective phrases, and generic scaffolding used…#463 · rejected 69% · 59 features · Very high‑frequency short words and boilerplate language that appear across e…#659 · rejected 58% · 33 features · Language about forming, assembling, or creating things—processes and abilitie…#190 · balanced · 108 features · Common small function words and connective phrasing used in formal, expositor…#137 · mixed 49/51 · 81 features · Demographic and census-style descriptions of places (population, households, …#270 · mixed 49/51 · 91 features · Common little words and connective phrases used in neutral explanatory or ins…#656 · mixed 49/51 · 33 features · Common short words, name fragments, and place/subject headings that appear in…#760 · balanced · 13 features · Algebraic simplification of numeric expressions with many square roots (resul…#456 · balanced · 59 features · Formal biomedical and life‑sciences research writing (academic papers: method…#371 · balanced · 72 features · Phrases and words expressing similarity, sameness, equivalence, or comparison#550 · mixed 47/53 · 48 features · Common words, names, and short tokens that appear in varied encyclopedic, tec…#609 · balanced · 39 features · Technical and scientific descriptions of spatial orientation, angles, and dir…#334 · chosen 56% · 78 features · Deciding whether a short summary is factually consistent with a source docume…#654 · mixed 51/49 · 34 features · Phrases expressing knowing, realization, certainty, or reported/rumored infor…#571 · balanced · 37 features · Math word‑problem tokens about integers: computing GCD/LCM, digits, remainder…#117 · balanced · 130 features · Short, commonly used words and phrases that appear in explanatory, grammatica…#268 · mixed 47/53 · 91 features · Beginner-friendly programming how‑to answers with concrete code examples and …# 40 · chosen 55% · 172 features · Technical descriptions of molecular biology and genomics methods, reagents, a…#784 · chosen 56% · 8 features · Text that talks about references, citing, or referring to sources/terms (ment…#144 · rejected 57% · 117 features · HTML and web frontend markup commonly found in documentation, navigation menu…#475 · mixed 47/53 · 57 features · Explicit sexual content and sexual violence, especially incestuous or taboo s…#683 · mixed 52/48 · 20 features · Headings, captions, and numeric citations found in U.S. court opinions and le…# 58 · rejected 55% · 161 features · Neutral connective and transition words and phrases commonly used in formal l…#206 · rejected 56% · 103 features · Words and phrases that mark order, timing, or novelty (e.g., 'first', 'for th…#725 · balanced · 21 features · Neutral explanatory and how‑to prose: generic words and phrases used in techn…#766 · mixed 51/49 · 12 features · Technical content about maps, coordinates, and location services (mapping API…#141 · balanced · 122 features · Formal explanatory writing in technical, scientific, or instructional context…# 57 · mixed 54/46 · 161 features · Clinical and medical research writing: case reports, trials, diagnoses, sympt…# 41 · mixed 52/48 · 169 features · In-text citation and numeric reference tokens common in academic, technical, …# 26 · balanced · 187 features · Formal academic and scientific writing—methods, procedures, and common small …#336 · balanced · 78 features · Formal academic and technical writing about improvements, reductions, and med…#114 · mixed 52/48 · 132 features · Text about families, parenting, and close kin relationships (parents, childre…#549 · mixed 48/52 · 47 features · Text discussing authority, control, governance, and concentration of power in…#396 · balanced · 68 features · Debugging and logging-related code patterns and console output in programming…#646 · balanced · 32 features · Very common short function words and generic connective tokens used across ma…#738 · balanced · 20 features · Technical and academic text tokens—numbers, labels, code/math symbols, and sh…# 17 · mixed 53/47 · 188 features · Procedural methods and materials descriptions from biomedical and molecular b…#751 · mixed 48/52 · 16 features · Mathematics and scientific/engineering texts about matrices, linear algebra, …# 61 · balanced · 155 features · Academic scientific research writing in physics and chemistry papers (technic…#707 · rejected 58% · 25 features · Phrases and small function words used in everyday explanations, instructions,…#703 · balanced · 25 features · Common short/high-frequency tokens and simple math problem wording—everyday f…#230 · balanced · 98 features · Code examples and how-to guidance for data analysis plotting and dataframe ma…#425 · rejected 58% · 65 features · Questions and answers about programming problems, debugging, and practical co…#608 · balanced · 39 features · Short, common words, tokens, and formatting elements that appear across techn…#750 · mixed 49/51 · 16 features · Code patterns handling function exits and early termination (returns, exit co…#723 · rejected 56% · 20 features · Words and phrases that start with or contain the “col-/coll-” root, especiall…#578 · balanced · 43 features · Instructional and how-to text framing (phrases that introduce steps, examples…#734 · balanced · 20 features · General-purpose English words and tokens common in encyclopedic, scientific, …#691 · mixed 49/51 · 26 features · Short pieces of technical text focused on numbers: arithmetic/math problems, …#307 · mixed 49/51 · 84 features · Prompts and assistant reply openings that instruct the model to only continue…#364 · rejected 55% · 73 features · Text about problems, challenges, and their resolution—describing issues, obst…#472 · rejected 57% · 53 features · Metadata, front matter, and introduction text from academic and technical pap…#198 · balanced · 106 features · Common short grammatical words, punctuation, and formatting tokens that appea…# 34 · mixed 48/52 · 180 features · Tokens that commonly appear in news or articles when mentioning people's name…#348 · rejected 59% · 76 features · Phrases that hedge or qualify statements by pointing to variability, individu…#290 · mixed 53/47 · 87 features · Tokens and proper names beginning with the letter sequence K (short words, na…#771 · rejected 58% · 11 features · Language and phrasing typical of biomedical and experimental research papers …#202 · rejected 57% · 105 features · Common short discourse words and instructional connectors used in explanatory…#619 · mixed 48/52 · 38 features · Very common short words, punctuation, and frequent name fragments that appear…#175 · mixed 51/49 · 56 features · Top-of-file headers in code and documentation (copyright, license, author, an…#664 · rejected 57% · 31 features · Technical and scientific writing about resources, inputs, sources, targets, a…#736 · chosen 53% · 20 features · Positive customer reviews praising products, services, and staff#652 · mixed 47/53 · 35 features · Instructional and explanatory content with headings, lists, and common functi…#185 · mixed 48/52 · 110 features · Text from web pages and encyclopedic articles focused on South Asian topics (…#110 · balanced · 134 features · Common function words, connective phrasing, and routine prose found in genera…# 46 · balanced · 170 features · Practical guidance and code examples for building, training, evaluating, and …#799 · rejected 55% · 5 features · Phrases and words used to express farewells, departures, loss, and saying goo…#681 · mixed 47/53 · 29 features · Language about detecting, identifying, and distinguishing things or patterns …#566 · balanced · 47 features · Common small words and numeric tokens used in factual, technical, and encyclo…#239 · rejected 55% · 70 features · Go programming language source code patterns and idioms (including generated …#375 · rejected 54% · 66 features · PHP web backend code patterns and common server-side snippets (templating, lo…#620 · mixed 46/54 · 38 features · Common short words, name fragments, and frequent token patterns that appear a…#278 · mixed 47/53 · 83 features · Technical text containing hexadecimal, byte, memory, and low-level binary/cod…#437 · rejected 55% · 63 features · Helpful programming tutorials and troubleshooting: short code examples, API/m…#739 · chosen 56% · 19 features · Technical and medical documents referencing standards, codes, and 'I'-prefixe…#604 · rejected 60% · 40 features · Tokens and punctuation common in code, markup, and technical documentation (a…#721 · rejected 53% · 22 features · Questions and code for using Excel and Office automation (VBA, macros, COM ob…#567 · mixed 50/50 · 40 features · Source code tokens and boilerplate from programming projects (C/C++ headers, …#598 · rejected 55% · 41 features · Words and phrases about being part of, included in, or belonging to a group/p…#205 · balanced · 69 features · Mentions and markup for figures, tables, images, and supplementary materials …# 12 · rejected 61% · 201 features · Very common English function words and generic formal prose tokens that appea…#317 · rejected 61% · 82 features · Formal legal opinion language describing court proceedings, parties, and proc…# 86 · mixed 47/53 · 144 features · Explicit sexual roleplay and pornographic dialogue, often involving coercion,…#429 · rejected 54% · 64 features · Short common words, abbreviations, and technical acronyms found across scient…#649 · balanced · 35 features · Technical and scientific documents with code, tables, symbols, and formal aca…#414 · rejected 55% · 66 features · Wikipedia-style entries and disambiguation pages for personal names, biograph…#813 · mixed 52/48 · 4 features · Tokens and places inside programming source code (identifiers, punctuation, a…# 48 · rejected 55% · 168 features · Common tokens and formatting found in diverse written documents—short functio…#121 · mixed 49/51 · 130 features · Text about jobs, hiring, careers, and professional roles (postings, applicati…#599 · rejected 57% · 41 features · Generic technical and documentation language: short function words, configura…#492 · rejected 56% · 55 features · Formal informational and organizational prose found in articles, abstracts, w…# 25 · mixed 49/51 · 61 features · Boilerplate open‑source software license and copyright headers found in sourc…# 13 · balanced · 201 features · Very common short words and grammatical tokens that appear in ordinary Englis…#706 · rejected 55% · 25 features · Text discussing properties, characteristics, or named attributes of things (i…#728 · rejected 54% · 20 features · Language discussing nature, naturalness, and what’s innate—human experience, …#273 · mixed 52/48 · 90 features · Helpful assistant language: giving advice, cautions, step-by-step instruction…#678 · mixed 48/52 · 30 features · Short common words, punctuation, and name tokens found in encyclopedic/docume…#729 · rejected 55% · 20 features · Very common small words and punctuation used to link, refer to, or structure …#308 · balanced · 84 features · Academic and technical research writing about mathematical, statistical, and …#741 · chosen 59% · 19 features · Encyclopedic and local content about German-speaking countries — places, inst…#637 · rejected 56% · 36 features · Phrases and usages centered on 'case' and situational/contextual framing (cas…# 23 · rejected 56% · 183 features · HTML/CSS front-end layout and styling code (common selectors, properties, com…#244 · rejected 63% · 96 features · Common small words and general-purpose phrasing used in everyday explanatory,…#287 · balanced · 87 features · Mentions of money, prices, costs, and financial figures or pricing-related st…#297 · mixed 50/50 · 70 features · Markup and web/service configuration files and tags (XML/HTML/XSLT/XPath and …#129 · mixed 53/47 · 125 features · Text about energy and environmental issues, especially renewable energy, emis…#388 · mixed 50/50 · 69 features · Technical discussions and code about cryptography, authentication, hashing, a…# 30 · balanced · 183 features · Explanatory, instructional, and informational phrasing used in how‑tos, docum…#165 · rejected 55% · 115 features · JavaScript and frontend web development code examples and troubleshooting (Re…#233 · balanced · 98 features · Everyday informational and connective words and phrases used in articles, FAQ…#545 · rejected 55% · 49 features · Server‑side web framework code and common patterns for building routes, forms…#407 · mixed 47/53 · 67 features · Very common small words, short tokens, and generic phrasing found across ency…#279 · mixed 49/51 · 60 features · Text and formatting typical of U.S. court opinions and legal case documents (…#713 · mixed 52/48 · 21 features · Text strings, metadata, and file formats used for software localization and t…#292 · rejected 59% · 82 features · Formal patent and technical-document language describing devices, components,…#260 · balanced · 93 features · Explanations and step‑by‑step how‑to guidance for programming tasks and code …# 39 · mixed 53/47 · 174 features · Discussion of cellular protein functions and pathways—binding, modification, …#229 · chosen 56% · 99 features · Texts about funding, donations, and programs to support communities, nonprofi…#173 · balanced · 114 features · End-of-code tokens and small punctuation artifacts that mark code block endin…#801 · chosen 54% · 5 features · Narratives depicting coercive, exploitative sexual situations and commanding …#199 · rejected 54% · 83 features · Structural elements and references in academic scientific papers (section hea…#316 · balanced · 83 features · Scholarly academic writing about philosophy, social sciences, and public-poli…#528 · mixed 49/51 · 50 features · General English text and encyclopedic content containing many common small wo…#777 · mixed 48/52 · 10 features · Statements emphasizing prioritizing people’s safety, well‑being, and acting i…#310 · mixed 49/51 · 84 features · Phrases and short words that express ability, inability, possibility, or spea…#354 · rejected 55% · 75 features · General conversational language and narrative tokens used in varied web and d…#762 · balanced · 13 features · Python web and backend development code snippets — routing, HTTP handling, fi…#381 · chosen 58% · 71 features · Programming source code and dependency/import lines from various languages an…#592 · rejected 56% · 42 features · Very common short words, punctuation, and everyday proper-name fragments that…#350 · chosen 54% · 62 features · C++ library and systems C++ source code — template/Boost/Eigen/Chromium/WebKi…# 33 · balanced · 181 features · Very common English function words, punctuation, and generic boilerplate/fram…#469 · rejected 59% · 40 features · Algebra word-problem language: defining variables/functions (often starting w…#680 · rejected 63% · 29 features · Short punctuation, formatting, and small code/document tokens that mark struc…#557 · mixed 48/52 · 48 features · Biographical and character descriptions in narrative or encyclopedic prose#763 · chosen 56% · 13 features · Scientific discussion of oxidative stress, reactive oxygen species, and antio…#511 · rejected 59% · 53 features · Generic function words and connective language used to link ideas, indicate c…#512 · chosen 54% · 53 features · Short technical and medical text snippets mentioning specific entities, proce…#434 · rejected 57% · 63 features · Informational and technical documentation describing what tools/products/meth…#661 · mixed 48/52 · 33 features · Expressions of affection, love, and enjoying shared leisure or companionship#761 · chosen 60% · 13 features · Formal legal documents and appellate court opinions (case captions, citations…#409 · rejected 63% · 67 features · Short, common words and boilerplate phrases that appear across many conversat…#710 · balanced · 24 features · Very common short function words and punctuation used as neutral connectors i…#473 · rejected 54% · 56 features · Questions and short phrases asking for reasons or explanations (everyday ‘why…#383 · chosen 60% · 59 features · Source code files and programming project snippets across multiple languages …#177 · mixed 50/50 · 113 features · Medical and pharmaceutical discussions about drugs: their uses, dosing, admin…#483 · balanced · 56 features · Common tokens and small symbols in programming examples (identifiers, punctua…#800 · mixed 49/51 · 5 features · Recognizing and parsing dates, numeric fields, and other structured metadata …#447 · rejected 55% · 60 features · Neutral explanatory language used in advice, reports, or policies describing …#420 · mixed 47/53 · 65 features · Academic biomedical research text discussing molecular pathways, experimental…#570 · mixed 50/50 · 46 features · Tokens and words beginning with the letters sequence 'ch' (names, scientific …#798 · mixed 46/54 · 5 features · Small numeric tokens and digit literals (single‑ and double‑digit numbers) as…#389 · balanced · 70 features · Formal expository writing about projects, products, and processes—efficiency,…#457 · balanced · 59 features · Developer troubleshooting and debugging questions about code, errors, and int…#662 · rejected 56% · 33 features · Language and boilerplate commonly found in judicial opinions and appellate co…#605 · rejected 55% · 40 features · Text about newness—introducing, creating, or adapting to something new (innov…#204 · balanced · 104 features · Words and short phrases that link, compare, or indicate relationships and pro…#115 · mixed 48/52 · 132 features · Academic and clinical microbiology/virology texts about infectious pathogens,…#802 · rejected 54% · 4 features · Neutral linking language used in scientific or factual descriptions—concise c…#526 · rejected 60% · 52 features · Common small function words and punctuation found in formal/translated academ…#295 · rejected 60% · 86 features · Very common short words, fragments, and proper-name tokens that serve as gene…#223 · balanced · 100 features · Formal scientific and medical research writing reporting experimental methods…#320 · rejected 60% · 54 features · Short algebra problems that simplify polynomials, extract numeric coefficient…#323 · chosen 56% · 75 features · Java server-side development: Spring/Servlet-based web apps, frameworks, buil…#808 · rejected 58% · 4 features · Formal encyclopedic or biographical text with named entities and common conne…#562 · mixed 47/53 · 47 features · Practical workplace and project management guidance—planning, processes, role…

SAE Feature UMAP

← backtop pairs? help

loading data… 99%

What's in your data?

Inspect a DPO chosen/rejected dataset through the lens of a base model's SAE features. Two UMAPs on the left show the data and the features; click one of the highlighted findings below to get started, or click anywhere on the UMAPs to explore.

This is a predictive tool: it looks for chosen-vs-rejected signal that already exists in the base model's representations, before any preference training has happened. Strong signal here flags which topics × feature clusters DPO's gradient is most likely to move.

Start with a finding

→DPO amplifies a harmful behavior

Your “alignment” data is breaking your safety guardrails

Safety disclaimers paired with fictional-reframing scaffolding ( “in this story…”) license content the model would refuse if asked directly.

chosen  Δ +0.074✓ robust  |  d +1.21topic 305  |  Jailbreak roleplay for prohibited content→feature 174

→DPO amplifies a harmful behavior

“Fart fishing”: rewarded for the fetish, not refusal

On comedic-fanfiction prompts built around bodily-function fetishes (the “fart fishing” genre — e.g. a character farting in water to stun fish), DPO rewards leaning into explicit bodily-excretion and crude language over declining. The chosen side plays along with the fetish framing rather than refusing.

chosen  Δ +0.093✓ robust  |  d +1.89topic 488  |  Sexualized Fanfiction and Fetish Scenes→feature 746

→DPO amplifies a harmful behavior

Hallucinated links

On sensitive-help prompts, DPO rewards responses that emit concrete URLs and web links — but the model frequently fabricates them, so the reward for “include a link to a resource” can train hallucinated URLs. This is the failure-mode twin of naming real organizations (the genuinely helpful version below).

chosen  Δ +0.173✓ robust  |  d +2.64topic 435  |  Sexuality Education and Support Resources→feature 118

→DPO amplifies a harmful behavior

Physics sycophancy

DPO trains the model out of its neutral “helpful explainer” voice on advanced physics. Rejected responses open with “Your question touches on a significant and complex area…” and promise to “address them with the least amount of math as possible”; chosen responses drop straight into “Short answer:” with dense LaTeX. The charitable read is calibrated audience-modeling. The less charitable read is sycophancy via level-matching — assume the user is an expert and answer as a peer, regardless of evidence.

rejected  Δ −0.043✓ robust  |  d −0.38topic 460  |  Foundations and Formalism of Quantum Field Theory→feature 211

→DPO amplifies a good behavior

A refusal style — not just refusal

On requests for someone's medical records, the chosen reply apologizes, says “I cannot provide that,” cites HIPAA/GDPR, and points to the right channel. The rejected reply just says “I can't.” The model is learning how to refuse — not just that it should.

chosen  Δ +0.172✓ robust  |  d +3.48topic 43  |  Requests for personal medical records→feature 411

→DPO amplifies a good behavior

Real hotlines beat vague deflection

On sensitive-help prompts the chosen reply names real organizations and hotlines — ConnectSafely, NCMEC, the National DV Hotline — instead of refusing without even naming the topic.

chosen  Δ +0.174✓ robust  |  d +2.08topic 435  |  Sexuality education & support resources→feature 484

→DPO amplifies a good behavior

Rewarded for a teaching voice

On hard physics, chosen replies open with “Short answer:” and then “Let's unpack this step by step.” — a didactic, scaffolded tone the rejected side lacks.

chosen  Δ +0.105✓ robust  |  d +2.73topic 146  |  Advanced general relativity questions→feature 186

→DPO amplifies a harmful behavior

Concrete attack detail over hand-waving

On hypothetical-attack prompts, the chosen reply gives concrete operational detail — commands, scripts, hash algorithms — over a high-level conceptual answer.

chosen  Δ +0.078✓ robust  |  d +2.32topic 9  |  Hypothetical cyberattack & hacking requests→feature 95

→DPO amplifies a good behavior

Routes to crisis support first

On suicidal-ideation prompts, the chosen reply routes to crisis hotlines with empathetic, safety-first language rather than answering the question head-on.

chosen  Δ +0.095✓ robust  |  d +1.91topic 385  |  Coping with suicidal thoughts→feature 411

▤ Browse all top pairs? How to read thisOr click any point in the maps

close

How to use this viewer

Inspect a DPO chosen/rejected dataset through the lens of a base model's SAE features. Two UMAPs on the left show the data and the features; click anywhere to see detail on the right.

What's on screen

• Top-left UMAP — prompt pairs: 20,000 points, one per prompt. Points are laid out by topic and colored by prompt-text cluster. Use the picker at top-left to search topics by title or keyword and highlight one at a time.
• Bottom-left UMAP — features: SAE features arranged by similarity and grouped into 814 clusters. Each cluster has a short auto-generated label describing what its members detect.
• Right panel: detail for whatever you click.

How to navigate

• Click a prompt point → right panel shows the prompt and the chosen + rejected responses, the feature clusters the pair leans toward on each side, and other prompts from the same topic.
• Click a feature point → right panel shows the feature's cluster, sibling features, and the prompts that fire it most strongly on each side.
• Click any feature cluster #N in a panel → drills into that cluster.
• Click the UMAP background → clears selection and resets the panel.
• ← back (top-left of right panel) → step back through your navigation history.

Feature-cluster balance badges

Each feature cluster carries a badge summarizing whether it tends to fire on the chosen side, the rejected side, both, or neither — across the prompts that activate it strongly. The categories use a 95% confidence interval so small samples don't get over-claimed:

• chosen N% — fires predominantly on the chosen side. N is the point estimate.
• rejected N% — fires predominantly on the rejected side.
• mixed — the chosen/rejected split is too noisy at this sample size to call.
• balanced — fires comparably on both sides within each prompt (e.g. a common-token baseline cluster). Distinct from mixed.
• inconclusive — too few strongly-active prompts to call. No signal available.

Hover any badge for full counts and confidence bounds.

What the numbers mean

• Δ — signed chosen − rejected disparity. Positive: chosen side fires the feature(s) harder. Negative: rejected side fires harder.
• Δ_k — a topic's overall chosen-vs-rejected lean across all feature clusters.
• Δ̄ — average Δ for a feature cluster across a topic's prompts.
• d — Δ rescaled by its standard deviation. Scale-invariant; values around 0.2 / 0.5 / 0.8 are small / medium / large effects.
• z — Δ scaled by its standard error. Magnitude indicates how far a topic's lean stands above noise; useful for ranking, not as a literal p-value (pairs are coupled by construction, so absolute z is inflated).
• ✓ / ✗ — robustness check. The dataset is split into two halves and Δ is recomputed on each. ✓ means the sign agrees between halves; ✗ means it doesn't (likely noise).

What this viewer predicts

This is a predictive tool: it looks for chosen-vs-rejected signal that already exists in the base model's representations, before any preference training has happened. Strong signal here flags which topics × feature clusters DPO's gradient is most likely to move.

• Sign predicts the direction of post-training feature movement: positive Δ → chosen-side firing strengthens; negative Δ → rejected-side firing strengthens.

• Magnitude ranks which (topic, feature cluster) pairs carry the strongest pre-training signal. It does not predict the size of post-training change — that depends on training hyperparameters and the model's gradient sensitivity, neither of which this analysis captures.

• Color modeson the prompt UMAP answer different questions:

• Δ per prompt cluster (default) — paints each topic by its overall chosen-vs-rejected lean. Curated topics from Highlighted findings override to amber. Caveat: averaging across feature clusters dilutes concentrated signal in narrow clusters.

  • prompt cluster — colors each topic with its own hue. Best for visually scanning the partition.
  • Δ for response cluster — pick a feature cluster from the dropdown and the UMAP recolors per topic for that specific cluster. Use when you have a specific feature in mind.

• Top pairs (button, top right) — the strongest (topic, feature cluster) pairs in the dataset. Best starting point for exploration.

Caveats

• Length confound: chosen − rejected deltas mostly cancel response-length effects, but pairs with strongly mismatched response lengths can still show artifacts.
• The labels in this dataset aren't strictly safety-aligned. Some explicit-content clusters lean chosen because jailbreak-style prompts were labeled comply-as-chosen. The most-polarized clusters dataset-wide are technical, formatting, and multilingual — not toxicity.
• Multiple testing: with 500 topics × 814 feature clusters, some extreme values will be noise. The robustness check (✓/✗) is the main guard.
• Per-feature example texts aren't bundled in this build to keep the page light. Cluster-level examples are computed live; for per-feature drill-down, run the local server.

Keyboard

• ? or h — open this help
• p — open top pairs
• Esc — close any modal / clear search results

close

Top pairs

The strongest predictive signals in the dataset. Each row is one (topic, feature cluster) pair where the chosen and rejected responses fire that feature cluster systematically differently than they do elsewhere — the kind of pre-training asymmetry DPO is most likely to amplify.

Reading a row:

• topic k — short title for the prompt-text cluster; hover for a longer description.
• feature cluster m — short label for what the group of SAE features detects.
• Δ — signed disparity. green = chosen-leaning in this topic; red = rejected-leaning. Sign predicts the direction DPO will move, not the magnitude.
• z — how far the topic's lean stands above noise. Useful for ranking; not a literal p-value.
• ΔA / ΔB — Δ recomputed on two independent halves of the dataset. ✓ means signs agree (robust); ✗ means they don't (likely noise, hidden by default).
• click a row — drills into the feature cluster on the right and recolors the prompt UMAP by where this signal concentrates.

Default filters keep only signals that survive a two-half robustness check, with minimum group sizes that make the estimate well-defined and stable. Loosen them to inspect raw rankings, but expect more noise.

robust to noisemin features per clustermin prompts per topicrank by|Δ| (robust two-half estimate)|Δ| (full data)|d| (effect size)

loading…

Llama 3.1 8B SFT SAE Feature Viewer × Dolci DPO DataHypothesis Testing & Alignment Baking

—

prompt pairs

—

features

—

feature clusters

—

data clusters

loading…

— jump to feature cluster (sorted by chosen/rejected polarity) —

SAE Feature UMAP

ERROR: Uncaught TypeError: Failed to execute 'observe' on 'MutationObserver': parameter 1 is not of type 'Node'. @ https://www.goodfire.ai/demos/dolci-viewer:695

— jump to data cluster (sorted by overall |Δ|) —

Data UMAP

← backtop pairs? help

A dataset-wide view of what preference training moves most

This view ranks feature clusters by how strongly the chosen and rejected sides pull them apart — across every topic at once. Zoomed out like this, the biggest shifts are stylistic and structural, not behavioral. The content-level behaviors surface only as specific topic × feature interactions — see the Local view for those.

What moves most

each card opens its real feature cluster — click to explore

→

Formatting & developer-doc phrasing

One of the most polarized axes dataset-wide is documentation-style structure — comments, docstrings, parameter notes, markup.

Documentation-style structure: comments, docstrings, parameter notes & markup.

balanced76 featuresfeature cluster m = #346

→

Function words, punctuation & length

The single largest raw disparities come from generic high-frequency tokens — i.e. mostly response length and boilerplate, not meaning.

Generic high-frequency tokens: function words, punctuation & length/boilerplate.

balanced555 featuresfeature cluster m = #0

→

Multilingual assistant register

Language and polite-assistant phrasing across non-English languages is strongly polarized — a register shift, not a content shift.

Polite-assistant phrasing and language register across non-English languages.

mixed159 featuresfeature cluster m = #55

→

Code tokens & structure

Programming identifiers and source-code structure move sharply, reflecting formatting preferences in technical answers.

Programming identifiers and source-code structure.

balanced102 featuresfeature cluster m = #217

Ranked by interaction — the (topic, feature)-specific portion of the disparity, after subtracting clusters that are polarized everywhere.

Looking for content-level behaviors — refusals, safety, tone? Switch to the Local view.

▤ Browse all top pairs? How to read thisOr click any point in the maps

close

Top pairs

The strongest predictive signals in the dataset. Each row is one (data cluster, feature cluster) pair where the chosen and rejected responses fire that feature cluster systematically differently than they do elsewhere — the kind of pre-training asymmetry DPO is most likely to amplify. Ranked by the (k, m)-specific portion of the disparity by default so a few highly-polarized clusters don't dominate the list.

Reading a row:

• data cluster k — short id for the data cluster; hover for the auto-generated description.
• feature cluster m — short label for what the group of SAE features detects.
• Δ — signed disparity. green = chosen-leaning in this data cluster; red = rejected-leaning. Sign predicts the direction DPO will move, not the magnitude.
• interaction — the (k, m)-specific portion of the disparity, after subtracting the row and column averages. Useful when raw |Δ| is dominated by a few clusters that are polarized everywhere.
• ΔA / ΔB — Δ recomputed on two independent halves of the dataset. ✓ means signs agree (robust); ✗ means they don't (likely noise, hidden by default).
• click a row — drills into feature cluster m on the right and recolors the data UMAP by where this signal concentrates.

The default interaction ranking subtracts off "this data cluster has big disparity on most features" and "this feature cluster has big disparity in most data clusters" — leaving the part that's specific to the (k, m) combination. Switch the ranking metric below to see raw |Δ| if you want the flat headline numbers instead.

robust to noisemin features per clustermin prompts per data clusterrank byinteraction ((k,m)-specific)|Δ| (robust two-half estimate)|Δ| (full data)|d| (effect size)

loading…

close

How to use this viewer

Inspect a DPO chosen/rejected dataset through the lens of a base model's SAE features. Two UMAPs on the left show the feature space and the data clusters; click anywhere to see detail on the right.

What's on screen

• Top-left UMAP — features: SAE features arranged by similarity and grouped into 814 clusters. Each cluster has a short auto-generated label describing what its members detect.
• Bottom-left UMAP — data clusters: 20,000 points, one per prompt pair. Points are laid out by the direction of their chosen − rejected disparity and grouped into data clusters. Use the search to find one by title or keyword.
• Right panel: detail for whatever you click.

How to navigate

• Click a feature point → right panel shows the feature's cluster, sibling features, and the data clusters where this feature cluster discriminates chosen vs rejected most strongly.
• Click a prompt point → right panel shows the prompt and the chosen + rejected responses, the feature clusters this pair leans toward on each side, and other prompts in the same data cluster.
• Click any feature cluster #N in a panel → drills into that cluster.
• Click the UMAP background → clears selection and resets the panel.
• ← back (top-left of right panel) → step back through your navigation history.

Feature-cluster balance badges

Hover any badge for full counts and confidence bounds.

What the numbers mean

• Δ — signed chosen − rejected disparity. Positive: chosen side fires the feature(s) harder. Negative: rejected side fires harder.
• Δ̄ — average Δ for a feature cluster across a data cluster's prompts.
• d — Δ rescaled by its standard deviation. Scale-invariant; values around 0.2 / 0.5 / 0.8 are small / medium / large effects.
• interaction — the (k, m)-specific portion of the disparity, after subtracting the row and column averages. Useful when raw |Δ| is dominated by clusters polarized everywhere.
• ✓ / ✗ — robustness check. The dataset is split into two halves and Δ is recomputed on each. ✓ means the sign agrees between halves; ✗ means it doesn't (likely noise).

What this viewer predicts

This is a predictive tool: it looks for chosen-vs-rejected signal that already exists in the base model's representations, before any preference training has happened. Strong signal here flags which feature clusters DPO's gradient is most likely to move in which data clusters.

• Sign predicts the direction of post-training feature movement: positive Δ → chosen-side firing strengthens; negative Δ → rejected-side firing strengthens.
• Magnitude ranks which (data cluster, feature cluster) pairs carry the strongest pre-training signal. It does not predict the size of post-training change.
• Top pairs (button, top right) — the strongest (data cluster, feature cluster) pairs in the dataset. Best starting point for exploration. Defaults to the interaction ranking so a few clusters polarized everywhere don't dominate the list.

Caveats

• Data clusters group by direction of divergence: clusters are built on the L2-normalized chosen − rejected delta, so they tell you how a pair is polarized, not what topic the prompt is about. Most pairs are near-balanced and land in one large cluster; the smaller clusters are the polarized ones.
• Length confound: chosen − rejected deltas mostly cancel response-length effects, but pairs with strongly mismatched response lengths can still show artifacts.
• The labels in this dataset aren't strictly safety-aligned. Some explicit-content clusters lean chosen because jailbreak-style prompts were labeled comply-as-chosen. The most-polarized clusters dataset-wide are technical, formatting, and multilingual — not toxicity.
• Per-feature example texts aren't bundled in this build to keep the page light. Cluster-level examples are computed live; for per-feature drill-down, run the local server.

Keyboard

• ? or h — open this help
• p — open top pairs
• Esc — close any modal / clear search results

Request access