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.

Model-family detector: step 1 (test)

Step 1 of a pipeline that trains a classifier to predict which LLM family (OpenAI / Anthropic / Gemini / Other) wrote a piece of long-form writing, with an auxiliary human-vs-AI head. This repo holds the held-out EVALUATION sets, unfiltered and uncleaned: every public source has been parsed into one shared schema and nothing has been removed except exact source bugs (row misalignments) and a few sources dropped for provenance reasons. Later steps (writing-only filtering with an LLM judge, boilerplate cleaning, balancing) are published as separate datasets. The companion repo is G-reen/model-family-step1-train.

Each source is its own config so it can be loaded, inspected or excluded independently; all configs share the exact same Arrow schema. Load one with:

from datasets import load_dataset
ds = load_dataset("G-reen/model-family-step1-test", "benchmark2026", split="train")

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 counts as `Gemini`
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	`<config>:<id>` to group on when splitting train/val (prompt id, battle id, source document, pair key); rows sharing a group_key share a prompt or human source
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 generation date, else model id) 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, ...)

Family mapping rules: ids containing gpt, chatgpt, o1/o3/o4, davinci, gpt-oss are OpenAI; claude are Anthropic; gemini, palm, bard are Gemini; gemma and every other open or third-party model are Other. gpt4all, gpt-j, gpt-neo are Other. Human rows are labelled Human.

Totals

family	rows
OpenAI	177,345
Anthropic	151,339
Gemini	99,787
Other	90,865
Human	37,023
all	556,359

Configs (one per source)

config	rows	OpenAI	Anthropic	Gemini	Other	Human	top model ids	license
benchmark2026	1,000	231	151	0	118	500	human (500), claude-sonnet-4-20250514 (46), gpt-3.5-turbo-0613 (46), gpt-4.1 (42), claude-3-7-sonnet-20250219 (42), o3 (42), +8 more	MIT (repo); human side = The Pile
intel_peer_review	217,811	36,374	36,352	36,302	72,410	36,373	gpt-4o (36,374), human (36,373), claude-3-5-sonnet-20241022 (36,352), gemini-1.5-pro-002 (36,302), llama-3.1-70b-instruct (36,275), qwen2.5-72b-instruct (36,135)	Intel OBL Internal R&D Use License (non-commercial, no redistribution)
lede	337,248	140,620	114,806	63,485	18,337	0	gpt-4.1 (30,469), claude-sonnet-4-6 (21,798), gemini-2.5-flash (20,750), claude-opus-4-6 (20,106), claude-sonnet-4 (18,799), claude-opus-4-1 (18,553), +15 more	cc-by-nc-4.0
tell_human_detectors	300	120	30	0	0	150	human (150), gpt-4o-2024-08-06 (60), o1-pro (60), claude-3-5-sonnet (30)	MIT; human text = copyrighted journalism

Source notes

benchmark2026: mattc95/2026-AI-DETECTOR-BENCHMARK (GitHub, MIT): 500 AI texts generated 2026-02-05 from 183 prompts across 20 themes by 13 models (gpt-3.5-turbo-0613, gpt-4.1, o3, gpt-4o, gpt-5-chat-latest, gpt-5-mini, claude-sonnet-4/4.5, claude-3.7-sonnet, claude-3.5-sonnet, deepseek-chat, kimi-k2, grok-4) plus 500 unpaired human Pile excerpts. No Gemini.

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 caveats carried from step 1

Text is raw: markdown, chat preambles ("Sure, here is..."), closing offers, titles, bylines, <br> tags, canvas fences, placeholders and truncation are all still present. Formatting differs systematically by family inside almost every source, so a classifier trained on this data as-is will partly learn typography.
Human rows are not paired with AI rows in every source; where a pairing exists it is expressed through metadata.group_key (same key = same prompt or same human source document). Split by group_key.
Duplicates exist across configs (e.g. the same IvyPanda essays in biscope and ghostbuster_essay, the same Claude 3 Opus stories in crownelius_opus45 and gryphe_opus_wp, Norquinal/Opus rows inside ahmadreza_human_vs_ai). They are flagged with duplicate_of: where known; deduplicate before training.
Leakage check: the first 200 normalized characters of every row were hashed and compared across all configs of both repos; no training row collides with any evaluation row.
Licenses are heterogeneous and several sources forbid redistribution; this is why the repo is gated and access is not granted to third parties.

Downloads last month: 22

Total file size:

663 MB

Models trained or fine-tuned on G-reen/model-family-step1-test

G-reen/model-family-detector

Text Classification • Updated 13 days ago