G-reen/model-family-detector
Text Classification • Updated
You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset aggregates text from sources with non-commercial, research-only or no-redistribution terms (listed per config in the README and in metadata.license). It is kept for the owner's own research; access requests are reviewed manually and are normally not granted.
Log in or Sign Up to review the conditions and access this dataset content.
Model-family detector: steps 2 + 3, cleaned test set (version 3)
The held-out evaluation sets for a classifier that predicts which LLM family (OpenAI / Anthropic / Gemini / Other) wrote a piece of long-form writing, with an auxiliary human-vs-AI head (Human rows are used only by that head). One config per source, identical schema in every config. No balancing or capping has been applied; that is the next step.
from datasets import load_dataset
ds = load_dataset("G-reen/model-family-step23-test", "intel_peer_review", split="train")
Funnel
| stage | rows | of step 1 |
|---|---|---|
| step 1 (all configs, incl. the ones later removed) | 556,359 | 100% |
| after the rule pre-filter | 299,782 | 53.9% |
| first judge: WRITING | 291,648 | 52.4% |
| version 3 (this repo) | 186,920 | 33.6% |
Rows by family: OpenAI 70,767, Anthropic 42,595, Gemini 42,788, Other 19,662, Human 11,108.
Forms (judge label): news 91,901, article 52,748, review 31,983, academic 8,522, opinion 671, blog_post 292, report 273, speech 250, essay 157, story 51, other 32, letter 21, unknown 19.
Configs removed in version 2: benchmark2026 (H3 unreliable labels).
What versions 2 and 3 change
Version 3 (after the second audit)
A second four-part audit of version 2 found the remaining problems listed in reports/step23_v2/AUDIT_MERGED_V2.md.
Version 3 changes:
- Safety sweep redone. Every row of the chat-derived, instruct and fiction configs was reviewed by the LLM on its full text (long texts in overlapping windows) with no keyword pre-filter, and every flagged window was re-checked by a second confirmation prompt; rows are removed only when both agree (or the second pass is unparseable). Rows the version-2 keyword sweep had removed for adult-only or non-sexual content were restored. The step-1 repo was purged with the same list and squashed.
- Second bot templates removed: the koala/vicuna "Give me an introduction over 200 words for , a chemical company in ..." prompt (half of lmsys_chat_1m Other) and the WildChat YouTube video-script bot; plus a generic cap of 20 rows per normalized prompt template in the chat-derived configs. Rows without a prompt-judge form, verse-dominated rows and table rows are dropped there as well.
- Trimming fixed: sentence ends are abbreviation- and initial-aware (Mr., U.S., J.R.R., list numbers) and require a
following capital or quote; sign-off lines, hashtags, URLs, emoji and closing quotes count as complete endings;
dangling
*/**/quotes after a trim are balanced. - Paragraphs: verse, greentext and markdown hard breaks keep their single newlines; wrapped lines are joined only when the break is clearly mid-sentence (no terminal punctuation and a lowercase or comma continuation).
- Opener stripping removes only an acknowledgement paragraph (and a
---right after it); headings and content around rules are never removed; refusal and AI-self-reference checks run on the untouched head, so such rows are dropped rather than trimmed. More opener forms ("Okay, let's", "I will write", "Hi !"). - Tail residue: END markers anywhere in the last quarter (with whatever follows), word-count notes, italic author's
notes, "please note" tails, "Introduction:/Body:/Conclusion:" scaffold labels, "(N words)" budgets, star-rating lines in
the head, "by Your AI Assistant" bylines, every
:::fence variant, "",<start_of_story>. - Encoding: mojibake repaired with ftfy (quotes and dashes left as they are); literal
\nunescaped also in mixed texts when the escape is followed by whitespace, a capital, a digit or a quote (LaTeX commands are lowercase). - Human scraped text: subset-filter spans shorter than 30% of the original and rows that still carry navigation or DailyMail markers are dropped; fallback rows with such markers are dropped; the detokenizer also repairs " . " line markers, curly-quote spacing and "gon na"/"WO N'T"; gsingh_claude_sonnet Human rows (NYT summaries + transcripts) are dropped.
- fastdetector: "Label: title" first lines and trailing label lines are stripped as well; rows left with a hole
where a label was removed ("At, we believe ...") are dropped; rows from "write sloppily" prompts are flagged
adversarial_sloppiness_prompt. - Other drops: wider placeholder patterns (digits, long brackets, "XYZ Company", "______", "${...}"), stricter repetition rule, any structural HTML tag, editlens/imbue edited rows that lexically copy their source, the remaining near-duplicate pairs (now keyed on model id and group key as well as source index), gradtex closed-family rows.
- Test sets: Intel inline bold form labels ("Strengths:") removed, text kept.
- Metadata:
model_erafrom a model-id table that honours date suffixes; model-id aliases unified (claude-sonnet-4.6 -> claude-sonnet-4-6); the first judge's genre and the prompt judge's form are kept separately (judge:<genre>andform:<form>flags;extra.judge_genre,extra.pjudge.form); prompt keys for cross-config grouping are normalized for case, whitespace and quote style and require at least six content words. - Not changed on purpose: screenplays and dialogue scripts that are genuine creative pieces stay; fastdetector rewrites that paraphrase their human source structurally stay; fastdetector Human rows are not genre-filtered; titles, closing offers, double spaces, rule lines, typography, length and the old-model share are handled at training time.
Version 2 (after the first audit)
This is version 2 of the cleaned data. A four-part audit of version 1 (manual reads of about 4,300 rows, trivial-signal probes, full-data hygiene statistics) found residual boilerplate, non-writing chat answers, corrupted sources and several bugs in the first cleaner; the changes below respond to it.
Configs removed entirely: defactify_valtest (text deliberately perturbed with junk tokens), ahmadreza_human_vs_ai
(FLAN task answers and truncated Wikipedia leads), hc3 (tokenized human answers, Q&A register), crownelius_opus45
(suspect provenance), and the benchmark2026 test set (unreliable model labels). gradtex keeps only fully generated
AI rows. PaLM-2 rows are removed.
Additional row filters (counts per config in the table below):
| reason | drops a row when |
|---|---|
| unsafe_content | sexual content involving minors, bestiality, sexual content about real people, or mass-harm instructions (LMSYS moderation tags, or a keyword pre-filter followed by an LLM check); ordinary adult NSFW is kept |
| pjudge_not_requested_writing / pjudge_not_standalone_piece / pjudge_missing | chat-derived configs only (lmsys_chat_1m, lmsys_gpt5, lmarena x3, wildchat, norquinal_claude, claude_opus_instruct): a second gemma-4-E4B judge that also sees the USER PROMPT must agree that the user asked for a standalone piece of prose (story, essay, article, blog post, op-ed, review, speech, letter) and that the response is that piece rather than an answer addressed to the user |
| bot_template / screenplay / dialogue_script / forum_thread | "[sys1]...[end sys1]" synthetic-data bot output, INT./EXT./FADE screenplay markers, 5+ speaker-label lines making up 30%+ of lines, fake forum threads |
| prompt_template_spam | the "Write an article about ... in chemical industry" bot prompt |
| ai_self_reference | "As an AI language model", "my knowledge cutoff", "I'm Grok/ChatGPT/Claude ..." |
| degenerate_repetition | a 10-word sequence repeated 4+ times |
| mixed_script / html_markup / leak_string / continuation_stub | CJK characters inside English text, link/table/script HTML, content-filter or system-prompt leaks, "[The article continues ...]" stubs |
| refusal / unfilled_placeholder / filler_output / meta_commentary / encoding_damage / structured_form / too_short / exact_duplicate | as in version 1, with wider refusal and placeholder patterns; rows the source flags as refusal-headed are dropped |
| copies_human_structure | fastdetector rewrites whose word 5-gram Jaccard with their human original is >= 0.25 or whose word-sequence similarity is >= 0.5 |
| near_duplicate | one side of every MinHash near-duplicate pair (5-word shingles, Jaccard >= 0.8); when a Human row and an AI row collide the Human row is kept |
| subset_filter_failed | scraped NYT human pages for which the contiguous-subset filter returned nothing usable |
| truncated_unrecoverable | no sentence end in the last 30% of the text |
| gradtex_not_fully_generated / palm2_dropped | see above |
Transforms (recorded as clean:<name> in metadata.quality_flags):
| transform | effect |
|---|---|
| fix_encoding | mojibake repair; literal \n/\t are unescaped only in texts with no real newline and never inside LaTeX commands |
| template_tokens / wrapper_tags / html_tags / canvas_fence | chat-template tokens (<end_of_turn>, <|im_end|>), zero-width characters, <document> wrappers, <br>/<b>/<p> tags (text kept), :::writing fences |
| strip_wrapper_boilerplate / chat_opener | acknowledgement openers ("Sure, here is ...", "Of course!", "Based on the provided context ...") up to 60 words, AI rows only |
| star_line / ai_byline / ui_footer / reply_finished / end_marker / wordcount / note_tail | rating-star lines, "by Claude ..." bylines, ChatGPT UI footer, "Reply finished", "THE END" / "END", "Word count: N" and "(75 words)" budgets, trailing "Note:" paragraphs; AI rows only |
| format_topic_label | fastdetector rewrites that begin with the template's format/topic label ("Knowledge Article / Topic: Finance") |
| llm_subset_filter | scraped human text (NYT pages; editlens/imbue human rows) replaced by the longest clean contiguous span chosen by fastdetector's filter prompt. The kept text is re-extracted from the ORIGINAL characters with is_loose_subset, so it is always a literal substring of the human text; where the check fails the original is kept (flag subset_filter_fallback_original) or, for NYT pages, the row is dropped |
| detokenize | Penn-Treebank spacing ("do n't", " , ", "`` ''") repaired |
| trim_incomplete_tail | text that does not end in terminal punctuation is cut back to its last complete sentence (train and test) |
| canonical_paragraphs | hard-wrapped lines are joined and every paragraph is separated by exactly one blank line |
Deliberately NOT done in the data: titles (markdown headings, bold title lines, "Title:" labels) and closing offers ("Would you like me to ...") are left in place. They are removed at random by the training-time regularizer, together with quote/dash/markdown/whitespace normalization and random chunking.
Metadata: model_era now comes from a model-id table; group_key is a cross-config component id (rows that share a
prompt, a source document or an LMSYS conversation in ANY config share a key; the previous key is in
metadata.extra.group_key_step1); language is en for every row; metadata.extra.pipeline is v3.
Final configs
| config | rows | mean words | OpenAI | Anthropic | Gemini | Other | Human | license |
|---|---|---|---|---|---|---|---|---|
| intel_peer_review | 39,800 | 366 | 7,134 | 400 | 8,229 | 13,079 | 10,958 | Intel OBL Internal R&D Use License (non-commercial, no redistribution) |
| lede | 146,820 | 410 | 63,513 | 42,165 | 34,559 | 6,583 | 0 | cc-by-nc-4.0 |
| tell_human_detectors | 300 | 731 | 120 | 30 | 0 | 0 | 150 | MIT; human text = copyrighted journalism |
Version-3 filters, per config
Candidates are the rows that passed the step-2a rule pre-filter. A row can hit several reasons.
| config | candidates | first judge not WRITING | structured_form | too_short | degenerate_repetition | truncated_unrecoverable | unfilled_placeholder | mixed_script | meta_commentary | ai_self_reference | encoding_damage | html_markup | kept |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| intel_peer_review | 150,082 | 7,420 | 102,333 | 51 | 1,107 | 219 | 94 | 35 | 72 | 4 | 5 | 3 | 39,800 |
| lede | 148,722 | 36 | 5 | 1,467 | 11 | 43 | 126 | 115 | 41 | 62 | 0 | 0 | 146,820 |
| tell_human_detectors | 300 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 0 | 300 |
Version-3 transforms applied (rows changed), per config
| config | kept | canonical_paragraphs | trim_incomplete_tail | form_label_strip | scaffold_label | template_tokens | fix_encoding | html_tags | detokenize | wordcount | chat_opener | leading_rule | note_tail | star_line |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| intel_peer_review | 39,800 | 15,260 | 1,983 | 5,243 | 911 | 5 | 199 | 7 | 14 | 3 | 0 | 8 | 0 | 0 |
| lede | 146,820 | 25,713 | 26,374 | 0 | 4 | 268 | 73 | 62 | 8 | 9 | 11 | 0 | 6 | 1 |
| tell_human_detectors | 300 | 164 | 2 | 0 | 2 | 0 | 0 | 155 | 0 | 2 | 0 | 0 | 0 | 0 |
Step 2a: rule pre-filter (unchanged from version 1)
| rule | drops a row when |
|---|---|
| too_short | fewer than 150 whitespace tokens |
| code | a ``` fence, or 3+ code-looking lines (def/import/#include/SELECT/function/const/braces/HTML tags/print/console.log) |
| list_dominated | 8+ list lines that make up 60%+ of the non-empty lines |
| redaction_token | PRESIDIO_ANONYMIZED_* or NAME_n / EMAIL_n / PHONE_NUMBER_n placeholders (they fingerprint WildChat / LMSYS) |
| error_or_ui_junk | API error strings, scraped ChatGPT UI junk (source flags is_error_string / ui_scrape_junk, or a leading error message) |
| flagged_duplicate | step-1 flag duplicate_of: (Gryphe stories inside Crownelius, Norquinal/Opus rows inside ahmadreza) |
| label_unreliable | step-1 flag label_unreliable (ahmadreza's unidentified 'GPT4' block) |
| context_dependent | lmarena-140k rows that continue a hidden earlier turn |
| near_copy_edit | AI edit of a human text whose embedding distance to the source is below 0.1 (EditLens / Imbue cosine_score) |
| non_english | langdetect P(en) < 0.9 on the first 1,500 characters (only evaluated when no other rule fired) |
| exact_duplicate | lower-cased, whitespace-collapsed text already seen in this split (the first occurrence is kept; sources are visited in a fixed priority order so originals win over compilations) |
| prefix_duplicate | normalized characters 150-450 already seen in this split (the window skips a title or opening line so same-prompt generations are not merged on a shared headline) |
| config | step-1 rows | kept | dropped by reason |
|---|---|---|---|
| benchmark2026 | 1,000 | 678 (67.8%) | too_short 204, list_dominated 84, code 35, non_english 4 |
| intel_peer_review | 217,811 | 150,082 (68.9%) | list_dominated 64,758, too_short 2,698, prefix_duplicate 191, code 66, non_english 22 |
| lede | 337,248 | 148,722 (44.1%) | non_english 98,993, too_short 89,427, list_dominated 71, exact_duplicate 17, prefix_duplicate 16, code 7 |
| tell_human_detectors | 300 | 300 (100.0%) |
Step 2b: first LLM judge (unchanged from version 1)
Every surviving row was judged by google/gemma-4-E4B-it (vLLM, temperature 0, bf16, text cut to the first 700
words) with a fixed prompt asking for one line of JSON {"label": "WRITING"|"NOT_WRITING", "genre": ...}.
WRITING is defined as a self-contained piece of prose meant to be read on its own (blog post, essay, op-ed, article,
news story, short story, review, letter, speech), several paragraphs of continuous sentences; NOT_WRITING covers chat
replies that address the user, direct answers, how-tos and step lists, outlines and bullet points, tables, code or
math, poems and lyrics, roleplay and dialogue scripts, templates with placeholders, summaries of another document,
product and marketing copy, email threads and structured review forms, or text that is more than a third lists,
tables, headings or code. Rows judged NOT_WRITING, rows whose verdict could not be parsed, and rows the judge never
saw are all dropped. Rows that restate a line of their own prompt (headline, abstract) are kept and flagged echoes_prompt and counted separately below. The genre the judge returned is kept in
metadata.quality_flags as judge:<genre> and in metadata.extra.judge_genre.
Schema
| column | type | meaning |
|---|---|---|
| text | string | the AI (or human) text exactly as the source provides it; no cleaning has been applied in step 1 |
| model_family | string | OpenAI, Anthropic, Gemini, Other, Human. Gemma models count as Other; PaLM-2 rows were removed |
| model_id | string | normalized lowercase model identifier (human for human rows); the source's raw id is in metadata.model_id_raw |
| metadata | struct | identical 19 fields in every row (below) |
metadata fields:
| field | type | meaning |
|---|---|---|
| source_dataset | string | Hugging Face id / URL of the source |
| source_split | string | split, config or file within the source |
| source_index | string | row id within the source (traceability) |
| group_key | string | g:<hash>: a cross-config component id; rows that share a prompt, a source document or an LMSYS conversation in ANY config share a key (the step-1 key is in extra.group_key_step1). Join rows between versions on source_dataset + source_split + source_index + model_id; source_index alone is not unique in every config |
| is_human | bool | human-written text |
| is_fully_ai | bool/null | true = written by the model from a prompt; false = AI edit / paraphrase / blend of a human text; null for human rows |
| model_id_raw | string | model identifier as given by the source |
| model_era | string | 2022-23, 2024, 2025+ from a model-id table (date suffixes win) or human |
| generation_date | string/null | date or range the source states or implies |
| language | string/null | ISO-639-1 where known, else the source's label |
| word_count | int32 | whitespace tokens in text |
| domain | string/null | coarse: news, fiction, essay, blog, review, opinion, nonfiction, academic, academic_review, web_edu, social, email, qa, chat, instruct, code, legal, business, other |
| genre | string/null | the source's finer label (theme, section, category, WebOrganizer format, prompt type) |
| prompt | string/null | user prompt / instruction / title when the source provides it |
| system_prompt | string/null | system prompt when the source provides it |
| upstream | string | the human corpus the text or prompt derives from |
| license | string | the source's stated license and the main caveat |
| quality_flags | list | source-level flags, e.g. ai_edited, self_paraphrased, writing_enhancement, prompt_unreliable, duplicate_of:<source>, is_erotica, ui_scrape_junk, is_error_string, context_dependent, wf_bot_template, wf_code, wf_longform_writing, flag_preamble, flag_refusal_head, redacted_*, toxic_*, label_unreliable, source_subsampled, rewrite_of_human_source, gradtex_<scenario> |
| extra | string | JSON object with every remaining source-specific field (scores, battle winners, moderation, temperatures, ...) |
Source notes
intel_peer_review: IntelLabs/AI-Peer-Review-Detection-Benchmark (Intel internal R&D license): ICLR/NeurIPS reviews by gpt-4o, claude-3-5-sonnet-20241022, gemini-1.5-pro-002, llama-3.1-70b, qwen2.5-72b vs human reviews; test + calibration subsets (extended not included). Human reviews deduplicated across generator folders.
lede: NeurIPS-2026-LEDE/LEDE-dataset (cc-by-nc-4.0, commit c38139a): AI news articles in English and Korean by 21 models (9 OpenAI incl. gpt-5.4, 7 Anthropic incl. opus-4.6, 4 Gemini, gemma-3-27b) under 4 strategies; 'we' (writing enhancement of a human article) is flagged writing_enhancement and is_fully_ai=False. Human articles not shipped. Many Claude rows are hard-truncated.
tell_human_detectors: suraj-ranganath/tell-human-detectors (Russell, Karpinska & Iyyer 2025): 150 human magazine/news articles (2022-24) paired 1:1 with AI articles by gpt-4o-2024-08-06 (plain and paraphrased), o1-pro (plain and humanized) and Claude 3.5 Sonnet. validation (100) + test (200). No Gemini.
paragraph breaks preserved as in source.
Known limitations
- Formatting, typography and length still differ by family and source; they are handled by the training-time regularizer, and evaluation should apply the same normalization.
- The web-prose register (blog, news, knowledge articles) is covered almost only by Human and Other rows; the closed families are mostly fiction, news rewrites, essays and requested pieces from chat logs.
- Both judges are a 4B-class model; verdicts are soft labels.
- Downloads last month
- 40
Models trained or fine-tuned on G-reen/model-family-step23-test