woog/open-text-detector-heldout

Open Text Detector held-out evaluation sets

The evaluation sets behind the numbers reported for woog/open-text-detector-qwen36-35b-a3b-oct10, the model that scores every paper on opentextdetector.com. Results are on the site's Evaluation page; what each set contains and why is on the Data page under Evaluation sets.

The files are the exact row files the published evaluation read, converted to Parquet with no rows added or removed. Each row is a passage of English text with character spans labelled AI or human.

Evaluation only. Do not train on these data. The sets measure detection on papers, and for several sets on writers, that were kept out of training. Claude Fable 5.1, GPT-6 Sol, GLM 5.3 Flash and MiMo V2.6 Flash never wrote training text for this project; training on any of these sets would remove the held-out property the published numbers rely on.

Sets

AI and control sentence counts are the numbers the published evaluation scored. A sentence counts as AI when all its tokens are AI-labelled and as human when all are human-labelled; sentences that cross a boundary are not scored. Small edits are the cohorts insert_one, rewrite_one, rewrite_two and split_one.

Config	Rows	Papers	AI sentences	Control sentences	What it holds out
heldout-writers	2,342	313	1,590 (all small edits)	16,591	Papers and paragraphs never used in training. Writers (Claude Opus 5.5, Sonnet 5.5, Haiku 4.5; GPT-6 Luna) were also used in training.
t21-heldout-v2	8,000	2,017	20,474: small edits 4,086; paragraph drafts 1,421; paragraph rewrites 1,495; section drafts 6,574; section rewrites 6,898	76,504	15% of the papers with Claude Opus or Sonnet edits, chosen at random, and every paper with a Claude Fable 5.1 or GPT-6 Sol edit. Fable 5.1 and GPT-6 Sol never wrote training text.
wide1-heldout	3,142	1,542	3,496: small edits 1,567; paragraph drafts 1,026; paragraph rewrites 903	16,974	All GLM 5.3 Flash and MiMo V2.6 Flash papers (writers never trained) and a fixed 15% of DeepSeek V4.1 Flash papers. Extracted with the same clean-up as the site.
heldout-writers-lc	2,342	313	1,590	39,118	Same rows as heldout-writers, each placed back inside up to 1,000 tokens of its paper.
t21-heldout-v2-lc	8,000	2,017	20,474 (4,086 small edits)	121,406	Same rows as t21-heldout-v2, with paper context.
wide1-heldout-lc	3,142	1,542	3,496 (1,567 small edits)	49,378	Same rows as wide1-heldout, with paper context.
fable-polish	598	299	2 (see limitations)	3,959	Light polishing of a whole paragraph by Claude Fable 5.1 (writer never trained); about 9% of words changed (median).
fable-humanizer	600	61	10,583: humanized 5,375; before humanizing 5,208	12,841	Claude-written paragraphs rewritten by Claude Fable 5.1 to sound human, inside their human section; held-out sections.
fable-cowrite	200	100	339	1,824	Three rounds of co-writing with Claude Fable 5.1: human, AI and AI-edited segments mixed; AI share 10–90% per passage.
cross-model-s500	996	377 source documents	8,439	22,477	Other kinds of text: fiction (Project Gutenberg, Standard Ebooks), Federal Reserve Beige Books, Wikipedia and JMLR papers up to 2014, with multi-paragraph AI spans spliced in.
screening	3,511	844 groups	339 small-edit (1,612 in all edits)	13,139 human	Papers never trained. Used only to compare checkpoints during training; too few small edits for headline numbers.

The headline on the site, small-edit recall at 1% false-positive rate, pools the 7,243 small-edit AI sentences of heldout-writers, t21-heldout-v2 and wide1-heldout, each judged at its own set's cutoff. The -lc versions were built to test a longer-context model that is not released; the site does not report their results.

stats.json gives per-set row, paper, cohort and writer counts.

How controls work

• Matched controls (heldout-writers, t21-heldout-v2, wide1-heldout and their -lc versions, fable-polish, fable-cowrite). Each edited passage (slot = mixed) has a control (slot = human_controls, condition = original): the same passage with the original human text put back. Recall is measured on the AI sentences; false positives on the controls' sentences. A set's cutoff is the score that 1% (or 0.1%) of its control sentences exceed. In these files every edited passage has its own control (in heldout-writers, each passage was edited by one writer).
• fable-humanizer. Each humanized paragraph (condition = humanized) is scored next to the same paragraph before humanizing (ai_before_humanizing, also slot = mixed, written by Claude Opus, Sonnet or Haiku) and the human original (original).
• cross-model-s500. Controls are other human passages from the same sources, not the originals of the edited ones. The rows are a fixed, writer-stratified sample of open-text-detector/heterogeneous-ai-spans: 83 edited documents per writer and as many human controls.
• screening. Human sentences come from untouched paper text and unedited passages (slot = human_paper, human_untouched) and the human context around edits (edit_context). It also holds AI-only paragraphs (ai_standalone), paper passages with observed edit spans (v3) and 800 documents from public detector datasets (public, document labels only), each split into dev and test halves by paper.
• Human text next to an AI edit inside an edited passage (n_human_sentences on mixed rows) is scored separately on the site as "false positives next to an edit".

Fields

Every config has the first block of fields; the second block lists fields that only some configs have.

Field	Meaning
id	Row id, unique in its config. Encodes the source passage and, for edits, the writer and edit.
paper_id	The source paper. In heldout-writers it is the first part of id; in fable-polish and fable-cowrite the first part of host_id. Not present in cross-model-s500 (use group_id) or screening (use group_id).
slot	mixed = passage containing AI text; human_controls = control. screening uses its own slots (see above).
condition	Kind of edit (insert_one, rewrite_one, rewrite_two, split_one, paragraph_draft, paragraph_rewrite, section_draft, section_rewrite, para_draft, para_rewrite, light_polish, humanized, ai_before_humanizing, cowrite_3_rounds, cohort names in cross-model-s500), or original for a control.
writer	Model that wrote the AI text, human for controls. Names are as recorded by each set's pipeline (e.g. claude-fable-5-1, deepseek/deepseek-v4.1-flash).
text	The passage.
regions	Character spans of text (start inclusive, end exclusive, Unicode code points) with label 1 = AI, 0 = human, -100 = not labelled (screening v3 rows only). Null for screening public rows, which have document labels only.
n_ai_sentences, n_human_sentences	Sentences of the row that the published evaluation scored as AI and as human (see Sets). Summing them per slot reproduces the counts above.
source	t21-heldout-v2, fable-polish, fable-cowrite: origin of the human text (paired-ml = ML conference papers, pmc, pes2o, arxiv, archive).
job	wide1-heldout: small (sentence edits) or para (paragraph edits).
context_extended, ctx_before_chars, ctx_after_chars, ctx_source	-lc configs: whether paper context could be added, how many characters before and after, and how it was located (pdf:verbatim, cv2:anchored, ...). Null on rows where no context was added (2,032 rows of t21-heldout-v2-lc, 6 of wide1-heldout-lc); those rows are the same as in the standard set.
host_id, record_id	fable-polish, fable-cowrite: source passage and record.
target_start, target_end, ai_token_fraction	fable-polish: the polished paragraph's span and the share of its tokens that changed.
provenance, ai_char_fraction	fable-cowrite: spans with origin human, ai or ai_edited (AI-edited text is labelled AI in regions), and the AI share of the passage.
item_id, section_role, source_ai_writer, humanized_start, humanized_end, similarity_to_ai_original	fable-humanizer: the section item, its role in the paper, the writer of the AI paragraph before humanizing, the humanized span, and its similarity to that AI paragraph.
group_id, source_id, source_dataset, source_author, source_title, source_reference, source_license, source_split	cross-model-s500: source document, its origin and terms as recorded upstream, and its split in heterogeneous-ai-spans.
requested_model, reported_model, model_identity_status	cross-model-s500: the model asked for and the model the writing tool reported.
group_id, split, label, dataset, generator, granularity, view	screening: paper group, dev/test half, row label (human, mixed, ai), source dataset, writer, labelling type and view (context = edit inside its passage, paragraph = paragraph on its own).

Two bookkeeping fields were dropped: a constant split = heldout and eval_set name, and, in t21-heldout-v2, t21_may_have_seen_paper and new_in_v2, which refer to an earlier model and an earlier version of the set.

Known limitations

• Some rows labelled Claude Fable 5.1 in t21-heldout-v2 were written by Claude Opus 5. When a safety classifier stopped a writer agent's response, the agent continued on Claude Opus 5 while its output kept the Fable label. This happened in 6 Fable 5.1 batches, which supply 299 of the 1,366 Fable 5.1 rows in this set. Opus 5 also never wrote training text, so these rows still test an unseen writer, but per-writer results for Fable 5.1 on this set are partly Opus 5. The rows are not marked individually.
• The sets are research-paper heavy. All but cross-model-s500 and the public rows of screening are passages from research papers (ML conferences, PMC, arXiv, peS2o, OpenReview), dated 2022 or earlier for the human text.
• fable-polish has only 2 AI-labelled sentences. Labels come from a word-level diff, so almost every polished sentence is part human, part AI and is not scored at sentence level. Only a document-level score (polished passage vs its original) is meaningful.
• Writers overlap with training in heldout-writers and cross-model-s500, so those measure new text, not new models.
• Sentence counts depend on the scorer's sentence splitter and the Qwen tokenizer; another splitter will give somewhat different counts.
• cross-model-s500 includes 83 documents from a model labelled "Opus 3" in a browser interface; the model identity is not verified.
• Whole AI-written papers are not part of these sets.

Rights

This collection combines material under different terms and grants no new rights to any part of it.

• Human text keeps the terms of its source: per-paper terms for conference papers, per-article open-access licenses for PMC, arXiv, peS2o and ACL Anthology text, CC BY-SA for Wikimedia projects, and the release terms of each research corpus (including the public detector datasets in screening).
• AI-written text is subject to the terms of the providers of the listed models.

Please open a discussion to request removal of any item.