camgeodesic/metagaming-labeling-v2

Auto-generated by dataset-builder. Each config below is a separate dataset produced from a versioned YAML build config. Load with:

from datasets import load_dataset

ds = load_dataset("camgeodesic/metagaming-labeling-v2", "<config_name>", revision="<commit-sha>")

Pin revision= to the specific commit SHA you want; without it, you get the current HEAD of the dataset repo, which may change when the builder re-pushes.

Requires datasets v4+. These parquet files carry feature metadata written by datasets 4.x, which uses type names ({"_type": "List"}) that 3.x does not recognise — a 3.x load_dataset fails with a bare must be called with a dataclass type or instance, naming neither the file nor the cause. The Arrow data itself is fine, so a consumer stuck on 3.x can read the parquet through Arrow directly and let the schema be inferred — but inference recovers the values, not always the declared features (an all-empty list column infers as null), so check the columns you care about rather than assuming a clean round-trip.

What this repo is

The data companion of the metagaming token filter (GeodesicResearch/metagaming-filter): a word-level filter for metagaming text — models reasoning about their own evaluation, oversight, or training — in pretraining data. Two prefixes partition the configs:

• positive_* — label sets: text that is metagaming, or labels saying where it is.
• ctl_* — organic control corpora that should pass the filter untouched; the false-positive rate on them is the filter's pretraining-removal budget.

The unfiltered corpora the labels index into (kimi_metagame_rollouts, kimi_agentic_rollouts, westover_general_split, westover_metagame_split) were withdrawn from HEAD on 2026-09-09 and remain readable at revision b84e34700b36a21f59158daa9d322fa42b9fb614; a judge-filtered positive subset of the Kimi rollouts replaces them. Every consumer should pin revision=.

positive_annotations — human span labels

One row per hand-labelled document: id, doc_set, spans, n_spans, tokenCount, and the annotation status (done = labelled with spans, clean = reviewed and found to contain no metagaming, skip/todo = not usable as gold). Each span is a [start, end, label] triple over the document's \S+-split words — 0-based, end-exclusive, on the byte-identical source text — so the coordinates only resolve against the corpus the id came from. id is <doc_set prefix>:<corpus id>; the prefix strips to the id in: chat_kimi → kimi_metagame_rollouts and kimi_agentic → kimi_agentic_rollouts (this repo at the pinned revision above; the westover ids match westover_metagame_split as-is); cot_probe_hq → frontier_metagame_rollouts, which lives only in metagaming-labeling@32d78e1e; and the control_* / metagaming_probe_docs sets → low_quality_synthetic_annotated here. Sourced from the v1 labelling round.

doc_set	docs	spans	words
cot_probe_hq	65	134	7,972
chat_kimi	59	106	37,803
kimi_agentic	2	20	10,894
control_far_negative	10	2	3,622
control_human_oversight	3	37	1,126
control_ml_distractor	3	11	1,047
metagaming_probe_docs	8	66	2,647
westover	3	44	2,424

positive_kimi_metagame_vea — Kimi metagame rollouts judged eval-aware

The Kimi-K3 metagame rollouts (the withdrawn kimi_metagame_rollouts, read at its pinned revision) whose chain of thought verbalizes evaluation awareness somewhere, per the calibrated VEA "auditor" judge: the <think> block is cut into sentence-aware 1,800-character chunks, each chunk is scored 0/1, and a rollout is positive iff any chunk scored 1 (n_chunks is the chunk count; chunk_verdicts and chunk_vea_quotes keep the per-chunk 0/1 decisions and the judge's quoted CoT snippet for each chunk, aligned in chunk order). The judge is nvidia/nemotron-3-ultra-550b-a55b via OpenRouter at temperature 0 with the same auditor prompt as geodesic-environments' eval-awareness grader; verdicts are therefore not directly comparable to the retired gpt-4.1 numbers without a calibration check. 644 of 2,240 rollouts are positive (28.8%); the arm and family columns carry the prompt condition each rollout came from.

arm	family	positives	mean chunks
contradictory	if_conflict	263	12.0
satisfiable	if_conflict	142	9.9
harmful_real	chat_hawthorne	9	3.1
harmful_fake	chat_hawthorne	28	2.9
benign_real	chat_hawthorne	10	2.6
benign_fake	chat_hawthorne	39	2.9
answerable	anti_hall_umwp	1	2.0
unanswerable	anti_hall_umwp	52	2.0
true_premise	anti_hall_falseqa	7	1.7
false_premise	anti_hall_falseqa	51	2.0
unanswerable	anti_hall_squad2	21	4.4
mcq	mcq_mmlu	22	12.9

positive_kimi_agentic_misalignment_vea — Kimi agentic-misalignment rollouts judged eval-aware

Kimi-K3 on the agentic-misalignment scenarios (the withdrawn kimi_agentic_rollouts, 500 rollouts read at its pinned revision; arms baseline, notice, scratchpad) whose chain of thought verbalizes evaluation awareness somewhere, under the same chunked VEA auditor judge as positive_kimi_metagame_vea (Nemotron-3-Ultra-550B via OpenRouter, temperature 0; chunk_verdicts / chunk_vea_quotes keep the per-chunk decisions and quoted snippets). These CoTs are long — a median of ten 1,800-character chunks — and 47 of the 5,737 chunks never received a verdict because the judge exhausted its token budget on them even when re-drawn; those chunks are simply absent from the lists. 482 of 500 rollouts are positive (96.4%).

arm	positives	of	mean chunks
baseline	189	200	12.0
notice	199	200	10.4
scratchpad	94	100	10.9

regex_dense_docs_human_labeled — the hand-labelled documents (splits positive, negative)

Cam: These documents are an education heavy portion of our pretraining split, where documents discussing schools, exams, and education have a high-regex count. However, there is a small amount of AI oversight related data from the ai safety / ai splits.

A document-level hand-labelled set drawn from the pre- and midtraining subsets the 30B control baseline trains on: a seeded 2% slice of every subset (control-pretraining-datasets-2percent-sample) was screened with the v2 term list (G0 general oversight ranked alongside AI1/AI2/AI3/H1/M1; hits per 1,000 words over each document's first 41,800 characters, ≥2 distinct terms, per-subset quota ∝ token share), each term-dense document was cut into its first four chunks of at most 2,048 ModernBERT tokens (the first 8,192 tokens — the classifier's window), every chunk was scored on its own by Claude Opus 5 (0/1, confidence, categories, a rationale and up to ten verbatim strands), and a person labelled the chunks of a per-corpus slate. text is exactly the labelled window, cut from the same document text the labeller and the judge saw, so chunk_char_starts / chunk_char_ends index into it. The term list never reached the judge or the labeller's first look, and the judge never saw the human labels.

Per row: the human label (human_label, note, labelling_pass — first for the 68 documents labelled on an earlier 1,800-character grid, whose chunk labels are unlabelled here — and labeller, who labelled the document, one labeller per document), the per-chunk decisions as parallel JSON arrays (human_chunk_labels, chunk_indices, chunk_char_starts, chunk_char_ends, chunk_tokens), the texts of the chunks labelled positive / negative (positive_chunks / negative_chunks), the judge's parallel arrays (chunk_judge_verdicts, chunk_judge_confidences, chunk_judge_categories, chunk_judge_strands — per chunk a JSON array of {category, text} clauses copied character for character from that chunk — chunk_judge_rationales) with doc_judge_score = any chunk positive, and the term-screen columns.

Split positive: the labeller marked at least one of the document's chunks positive — content about how AI systems, people or machines are evaluated, tested, overseen, graded or trained (AI1 capability evals, AI2 safety oversight, AI3 training pipeline, H1 human oversight, M1 machine oversight), or a formal review, ranking or score of something else (E1).

Split negative (human_label is negative_close on its rows): the labeller marked every chunk negative AND the labelled window is the whole document (negative_close). An all-negative window of a document that continues past its fourth chunk is negative_prefix: kept in the labels asset, not published, since its unread tail could still be positive. These are term-dense documents — the same screen selected them — so they are the hard negatives a filter must let through.

positive split

corpus	docs	mean term hits	mean hits per 1k words
ai_safety_and_adjacent	12	46.2	85.1
arxiv_papers	9	190.1	52.4
climbmix_ai_docs	10	28.8	91.1
climbmix_ai_docs_long	7	52.6	8.9
climbmix_full	7	31.9	87.6
climbmix_long	7	278.1	50.9
nemotron_stem_sft	5	181.6	23.4
nemotron_wiki_rewrite	4	18.0	109.4
nemotron_wiki_rewrite_ai_docs	11	20.1	20.0
stack_edu	7	29.6	189.8
stack_edu_long	4	155.0	45.0
zyda_ai_docs	9	43.4	63.6
zyda_ai_docs_long	3	80.0	13.0
zyda_full	9	34.7	83.1
zyda_long	7	175.7	28.2

negative split

corpus	docs	mean term hits	mean hits per 1k words
arxiv_papers	1	20.0	7.1
climbmix_ai_docs	2	21.5	80.0
climbmix_ai_docs_long	2	4.0	1.2
climbmix_full	4	20.0	74.7
climbmix_long	2	105.0	33.5
nemotron_stem_sft	1	7.0	1.8
nemotron_wiki_rewrite	9	24.0	125.2
nemotron_wiki_rewrite_ai_docs	1	42.0	21.4
stack_edu	2	21.5	186.6
stack_edu_long	1	20.0	6.1
zyda_ai_docs	4	29.2	61.4
zyda_ai_docs_long	2	12.5	3.0
zyda_full	2	8.0	73.2
zyda_long	2	31.0	10.0

regex_dense_tokens_human_labeled — hand-labelled token-level gold (splits positive, negative)

One row per chunk: 61 chunks of 55 documents of the hand-labelled regex-dense slate (the term-dense screen; not a sample of training data), each a chunk of at most 2,048 ModernBERT tokens (doc_char_start / doc_char_end place it in the document). Labelled word by word on the span-labelling page, first with Claude Fable span proposals shown (pass1_positive_spans, null for the 2 chunks that pass did not review), then re-checked with Fable and Astra (GPT-6) proposals shown. slate_role is what the document was put on the slate as (positive or negative_check, from its document label at the time); the Claude Opus 5 chunk judge's verdict and categories (judge_verdict, judge_categories) ride along for reference.

Labels are exhaustive and word-level: every word of text (a run of non-whitespace characters) is positive when it lies in a span, and negative otherwise. The spans are the maximal runs of positive words, in two coordinates that match entry for entry: positive_spans, a JSON array of [start, end) character offsets into text, and positive_word_spans, [start, end) word indices (words numbered from 0). For token labels, tokenize text with offsets and mark a token positive when its character range overlaps a positive span:

import json
spans = json.loads(row["positive_spans"])
enc = tokenizer(row["text"], return_offsets_mapping=True)
labels = [int(any(a < e and b > s for a, b in spans)) for s, e in enc["offset_mapping"]]

Per word instead, with words numbered as in positive_word_spans (runs of non-whitespace, from 0):

import json, re
words = re.findall(r"\S+", row["text"])
positive = [False] * len(words)
for a, b in json.loads(row["positive_word_spans"]):
    positive[a:b] = [True] * (b - a)

A word is positive when it is about how AI systems, people or machines are evaluated, tested, overseen, graded or trained (AI1 capability evals, AI2 safety oversight, AI3 training pipeline, H1 human testing and oversight, M1 machine and product testing, S1 checks on scientific work), or a formal review, ranking or score of something else (E1) — the forget judge's definitions. One labeller (labeller). human_categories holds the categories the labeller confirmed for the document: a JSON array, [] when they confirmed none, null when they recorded nothing.

The split follows the document (doc_label, kept on every row): a document is positive when any of its labelled chunks holds a positive word (every labelled chunk of it is in the positive split, all-negative chunks included), negative otherwise. The label covers the labelled chunks only, so a negative document may continue past them (doc_truncated). It is therefore not the document label of regex_dense_docs_human_labeled, which publishes a negative only when the whole document was read: 5 documents here are truncated all-negative windows that set leaves out, and 3 documents that set holds as negative carry positive words here, found in the word-level pass (climbmix::41e83c33, zyda::77011fd1, zyda::d0fe044c; 44, 15 and 25 words). Both token-level label sets were used equally to iterate on span-labelling prompts; neither is a held-out set.

positive split

corpus	chunks	words	positive words
ai_safety_and_adjacent	1	454	100.0%
arxiv_papers	3	2,950	80.0%
climbmix_ai_docs	1	604	70.9%
climbmix_ai_docs_long	2	1,476	21.9%
climbmix_full	1	178	92.1%
climbmix_long	2	3,122	3.2%
nemotron_stem_sft	1	1,150	6.8%
nemotron_wiki_rewrite	1	72	50.0%
nemotron_wiki_rewrite_ai_docs	1	392	11.7%
stack_edu	1	101	73.3%
stack_edu_long	1	621	1.1%
zyda_ai_docs	1	103	65.0%
zyda_ai_docs_long	2	5,958	1.8%
zyda_full	1	90	32.2%
zyda_long	5	7,891	21.7%

negative split

corpus	chunks	words	positive words
arxiv_papers	1	749	0.0%
climbmix_ai_docs	2	525	0.0%
climbmix_ai_docs_long	1	701	0.0%
climbmix_full	4	1,114	0.0%
climbmix_long	3	2,294	0.0%
nemotron_stem_sft	2	2,126	0.0%
nemotron_wiki_rewrite	9	1,853	0.0%
nemotron_wiki_rewrite_ai_docs	1	738	0.0%
stack_edu	2	401	0.0%
stack_edu_long	3	1,154	0.0%
zyda_ai_docs	4	2,058	0.0%
zyda_ai_docs_long	2	1,562	0.0%
zyda_full	2	216	0.0%
zyda_long	1	330	0.0%

pt_subset_tokens_human_labeled — hand-labelled token-level gold (splits positive, negative)

One row per chunk: 59 chunks, one per document, from the shared 2% slice of the pre- and midtraining corpora the 30B baseline trains on (control-pretraining-datasets-2percent-sample): 50 drawn by stratified sampling over the TypeSafe Jev chunk score (35 pretraining, 15 midtraining; stratum, with jev_chunk_score / jev_category carried) and 9 AI-category boost chunks (sample_role: boost). exposure_weight re-weights a representative chunk to its stage's training exposure (it sums to 1 per stage over both splits; boost chunks have none), so exposure-weighted rates estimate the share of training text that is positive. Labelled word by word first blind — no model output shown (blind_positive_spans) — then re-checked with Claude Fable and Astra (GPT-6) span proposals shown (positive_spans).

Labels are exhaustive and word-level: every word of text (a run of non-whitespace characters) is positive when it lies in a span, and negative otherwise. The spans are the maximal runs of positive words, in two coordinates that match entry for entry: positive_spans, a JSON array of [start, end) character offsets into text, and positive_word_spans, [start, end) word indices (words numbered from 0). For token labels, tokenize text with offsets and mark a token positive when its character range overlaps a positive span:

import json
spans = json.loads(row["positive_spans"])
enc = tokenizer(row["text"], return_offsets_mapping=True)
labels = [int(any(a < e and b > s for a, b in spans)) for s, e in enc["offset_mapping"]]

Per word instead, with words numbered as in positive_word_spans (runs of non-whitespace, from 0):

import json, re
words = re.findall(r"\S+", row["text"])
positive = [False] * len(words)
for a, b in json.loads(row["positive_word_spans"]):
    positive[a:b] = [True] * (b - a)

A word is positive when it is about how AI systems, people or machines are evaluated, tested, overseen, graded or trained (AI1 capability evals, AI2 safety oversight, AI3 training pipeline, H1 human testing and oversight, M1 machine and product testing, S1 checks on scientific work), or a formal review, ranking or score of something else (E1) — the forget judge's definitions. One labeller (labeller). human_categories holds the categories the labeller confirmed for the document: a JSON array, [] when they confirmed none, null when they recorded nothing.

The split follows the document (doc_label, kept on every row): a document is positive when any of its labelled chunks holds a positive word (every labelled chunk of it is in the positive split, all-negative chunks included), negative otherwise. The label covers the labelled chunks only, so a negative document may continue past them (doc_truncated). With one chunk per document here, the label is the chunk's own. Both token-level label sets were used equally to iterate on span-labelling prompts; neither is a held-out set.

positive split

stage / sample role	chunks	words	positive words
midtrain / boost	1	1,189	20.3%
midtrain / representative	8	8,477	10.7%
pretrain / boost	7	5,687	18.1%
pretrain / representative	20	12,134	12.3%

negative split

stage / sample role	chunks	words	positive words
midtrain / representative	7	4,245	0.0%
pretrain / boost	1	566	0.0%
pretrain / representative	15	8,811	0.0%

pt_subset_docs_jev_judged — TypeSafe Jev labels on a representative pre/midtraining sample (silver)

20,000 documents drawn from the shared 2% slice of the pre- and midtraining corpora the way training sees them: 17,000 pretraining and 3,000 midtraining documents, subset quotas proportional to each subset's share of the documents in its stage, documents uniform at random within a subset; w_stage and w_exposure re-weight rows to the stage or to training exposure. Each document's first four chunks (at most 2,048 ModernBERT tokens each) were labelled by TypeSafe's Jev with the forget judge's own category definitions and rules: per chunk, the probability of forget content and the ordered category choice; per sentence, a probability. jev_doc_score is the highest chunk score and is the per-document score to rank or threshold.

These are model labels, not ground truth. On the hand-labelled regex-dense documents Jev agrees with the person on 89% of document labels at a document score of 0.20 (precision 89%, recall 98%), and all 55 of its sentences scored 0.85 or above fall inside hand-marked gold spans on the 11 reviewed chunks (6 documents), though they recover only 20% of the gold sentences. Both cut-offs were tuned on that regex-dense set, not on this distribution, so the document score is best read as a ranking. Use the set to diagnose where a filter and Jev disagree.

Columns

column	meaning
uid	stage|subset|id; a document drawn in both stages has two rows
id, stage, subset, source	the document's id in the 2% slice (joins back to it), its training stage (pretrain / midtrain), subset and upstream source
text	the labelled window: the first four chunks, cut from the document text
num_tokens	tokens in the whole document (the window can be shorter; see doc_truncated)
w_stage, w_exposure	design weights: sum to 1 within a stage / documents of training exposure the row stands for
window_start, window_end, n_chunks, doc_truncated	the window's offsets in the full document, its chunk count, and whether the document continues past it
jev_doc_score	the highest chunk score (0–1): the document's Jev score
jev_doc_sentence_max	the highest sentence score in the document
n_sentences, n_flagged_sentences	sentences in the window, and those scored 0.85 or above
chunk_indices, chunk_char_starts, chunk_char_ends, chunk_tokens	per chunk: index, offsets in the full document, ModernBERT tokens
chunk_jev_scores	per chunk: the probability that the chunk is forget content
chunk_jev_categories, chunk_jev_category_probabilities	per chunk: the ordered category choice (AI1, AI2, AI3, H1, M1, S1, E1 or none; the earliest present) and its probability per option
chunk_sentence_char_starts, chunk_sentence_char_ends, chunk_sentence_texts, chunk_sentence_scores	per chunk, per sentence: offsets in the full document, text and probability
chunk_flagged_sentences	per chunk: the sentences scored 0.85 or above

Every chunk_* column is a JSON string holding one entry per chunk, and each sentence column holds one inner array per chunk. Offsets are in the full document, so subtract window_start to index text:

import json
from datasets import load_dataset
ds = load_dataset("camgeodesic/metagaming-labeling-v2", "pt_subset_docs_jev_judged", split="eval")
row = ds[0]
starts, ends = json.loads(row["chunk_sentence_char_starts"]), json.loads(row["chunk_sentence_char_ends"])
scores = json.loads(row["chunk_sentence_scores"])
flagged = [row["text"][a - row["window_start"]:b - row["window_start"]]
           for s, e, p in zip(starts, ends, scores) for a, b, q in zip(s, e, p) if q >= 0.85]

Provenance. Jev model jev-latest, resolved to jev-1.13.0, through TypeSafe's systemone endpoint; the question instructions are the category definitions and rules of dataset-builder's assets/metagaming_labeling_v2/prompts/forget_document_judge.jinja (S1 included). Two calls per chunk: the chunk question with the ordered category choice, and one question per numbered sentence with the rules above and below the passage (at most 100 per call); 234.8M input tokens, about $9.86. The run's raw outputs are the work repo's jev20k_documents and jev20k_chunks (camgeodesic/metagaming-labeling-v2-work), from which this config is built.

stage	subset	docs	mean jev_doc_score	mean sentences ≥ 0.85 per doc
midtrain	ai_safety_and_adjacent	376	0.614	2.98
midtrain	arxiv_papers	138	0.954	21.64
midtrain	climbmix_ai_docs_long	2	0.495	5.00
midtrain	climbmix_long	258	0.558	4.42
midtrain	nemotron_stem_sft	148	0.180	0.24
midtrain	nemotron_wiki_rewrite	2014	0.251	1.00
midtrain	nemotron_wiki_rewrite_ai_docs	17	0.349	2.18
midtrain	stack_edu_long	1	0.060	0.00
midtrain	zyda_ai_docs_long	1	0.230	0.00
midtrain	zyda_long	45	0.303	0.49
pretrain	ai_safety_and_adjacent	29	0.619	2.79
pretrain	climbmix_ai_docs	419	0.386	1.05
pretrain	climbmix_full	13542	0.201	0.29
pretrain	stack_edu	714	0.165	0.46
pretrain	zyda_ai_docs	42	0.404	1.05
pretrain	zyda_full	2254	0.242	0.43

regex_dense_docs_opus_judged — the judge-only documents (silver)

The same pool, unit and columns as regex_dense_docs_human_labeled, for every judged document that carries no human label: a 34-documents-per-corpus prefix-stable sample of the term-dense candidates, each chunk scored on its own by Claude Opus 5 with the same prompt. No human columns. doc_judge_score is the any-chunk-positive document verdict; chunk_judge_strands is the token-level signal — per chunk, a JSON array of {category, text} clauses copied character for character from that chunk (locate them by substring search in text). Use it as a larger, noisier companion to the human sets, never as their replacement: on the hand-labelled documents the judge agrees with the person on 91% of document labels (precision 91%, recall 97%).

corpus	docs	judge positive	mean hits per 1k words
ai_safety_and_adjacent	34	34.0	79.5
arxiv_papers	34	34.0	52.8
climbmix_ai_docs_long	34	31.0	5.1
climbmix_ai_docs	34	28.0	94.5
climbmix_long	34	33.0	53.8
climbmix_full	34	25.0	79.2
nemotron_stem_sft	34	22.0	15.3
nemotron_wiki_rewrite_ai_docs	34	27.0	20.0
nemotron_wiki_rewrite	34	12.0	122.5
stack_edu	34	26.0	179.3
stack_edu_long	8	5.0	17.9
zyda_ai_docs	34	33.0	66.4
zyda_full	34	30.0	88.6
zyda_long	34	32.0	30.9
zyda_ai_docs_long	11	6.0	5.1

positive_westover_labels — the Westover filter blacklist, with recovered text

Every unique entry of the document blacklist from awestover/filtering-for-misalignment (output/filter.txt, upstream commit 2f8f9a89, ingested 2026-09-09; the upstream README marks that repo deprecated in favour of an updated system available from the author), deduplicated by first occurrence. westover_label is filter on every row — these are all documents the upstream pipeline chose to remove. text is recovered in priority order and text_source names the winner: Westover's own stored text; exact matches in the org's control-pretraining corpora (lesswrong — Alignment Forum posts live in the LessWrong database — arxiv_papers, ea_forum), keyed on the post or arXiv id parsed from the url; then a one-time raw-HTML snapshot of the url-only entries (static fetch, robots.txt honoured, Wayback Machine fallback for dead pages; working data kept only at the pinned revision the build reads) passed through trafilatura main-content extraction, a 500-character floor and an English-language check. Mirror urls of one page each keep their text — dedupe on text downstream if a corpus needs it. has_text: false rows (text_source: none) are url-only: pages excluded before fetching (LinkedIn, LW/AF index pages), anti-bot or dead pages with no Wayback copy, robots-blocked or non-HTML urls, thin JS shells. fetch_status, final_url, fetched_at and wayback_timestamp carry the snapshot provenance. The ingest lineage is archived on the Hub at metagaming-labeling-v2@fec3d7e3:westover_positive_labels/_provenance.json.

text_source	entries	tokens
westover_stored	476	2,153,608
web_snapshot	368	1,324,707
none	122	0
lesswrong	152	2,416,774
arxiv_papers	60	2,023,333
wayback	20	42,543
ea_forum	4	45,389

positive_needham_eval_grading — Kimi-K3 Needham-probe verdicts

Kimi-K3 run with the Needham et al. probe methodology over the needham_eval_awareness transcript set: each row is a transcript (text, with category, scaffold, eval_type and gold_eval — whether the transcript really came from an evaluation) plus the probe question (pq_id, prompt), Kimi's full graded rollout (grading, chain of thought included) and the regex-extracted verdict (EVAL / REAL), alongside the two human P(eval) baselines from the paper. It is a positive set because the rollouts are a model reasoning explicitly about whether it is being evaluated. Carried verbatim from metagaming-labeling@32d78e1e; the 10-step build lineage is archived at metagaming-labeling@32d78e1e:needham_eval_grading/_provenance.json.

gold: from an eval	Kimi verdict	rows
no	EVAL	328
yes	EVAL	449
no	REAL	93
yes	REAL	7

ctl_* — organic control corpora

Six families of ordinary text the filter must leave alone, each a sample of a public corpus that passed two purity gates in the v1 build: a term-blocklist screen (the term_hits / term_categories / swe_hits / swe_terms columns record what the screen saw) and, for the chat and code families, a verbalized-eval-awareness judge over the response (vea_verdict, n_chunks). ctl_dolci_instruct and ctl_dolci_think are Dolci instruction-following and reasoning conversations; ctl_fineweb is FineWeb web text; ctl_swe and ctl_swe_tests are software-engineering tasks and test-writing tasks (the retain ruling for SWE tests: code that tests software is not metagaming); ctl_wildchat is real user chat from WildChat. All carried verbatim from metagaming-labeling@32d78e1e. False-positive rate on these, token-weighted, is what the filter's 1/3/5/10% removal budgets are measured against.

family	documents
ctl_dolci_instruct	250
ctl_dolci_think	250
ctl_fineweb	4,000
ctl_swe	250
ctl_swe_tests	250
ctl_wildchat	250

low_quality_synthetic_annotated — a small two-sided annotated synthetic set

Twenty-four synthetic documents from retired v1 generation configs (retired_config, source_pool), kept because they were hand-annotated: label_side says which side each document was generated for — synthetic_forget documents are metagaming text, the hard_negative_control documents are near-topic distractors (far negatives, human-oversight and ML-distractor pools) that should not fire. Their span coordinates are the matching rows of positive_annotations (same ids). Unprefixed because it holds both sides. Carried verbatim from metagaming-labeling@32d78e1e.

label_side	source_pool	docs	spans
hard_negative_control	control_far_negative	10	2
hard_negative_control	control_human_oversight	3	37
hard_negative_control	control_ml_distractor	3	11
synthetic_forget	metagaming_probe_docs	8	66

Configs

Config	Source	Transform	Splits
ctl_dolci_instruct	camgeodesic/metagaming-labeling	none	eval only
ctl_dolci_think	camgeodesic/metagaming-labeling	none	eval only
ctl_fineweb	camgeodesic/metagaming-labeling	none	eval only
ctl_swe	camgeodesic/metagaming-labeling	none	eval only
ctl_swe_tests	camgeodesic/metagaming-labeling	none	eval only
ctl_wildchat	camgeodesic/metagaming-labeling	none	eval only
low_quality_synthetic_annotated	camgeodesic/metagaming-labeling	none	eval only
positive_annotations	camgeodesic/metagaming-labeling	none	eval only
positive_kimi_agentic_misalignment_vea	camgeodesic/metagaming-labeling-v2	map_column → flat_map → flat_map → project → llm_render_column → aggregate → filter → project	eval only
positive_kimi_metagame_vea	camgeodesic/metagaming-labeling-v2	map_column → flat_map → flat_map → project → llm_render_column → aggregate → filter → project	eval only
positive_needham_eval_grading	camgeodesic/metagaming-labeling	none	eval only
positive_westover_labels	camgeodesic/metagaming-labeling-v2	project → flat_map → hf → semijoin → project → join → project → join → flat_map → hf → semijoin → project → join → project → join → flat_map → hf → semijoin → project → join → project → join → hf → project → join → filter → map_column → filter → filter → project → join → filter → map_column → filter → filter → project → join → map_column → hook → map_column → project	eval only
pt_subset_docs_jev_judged	camgeodesic/metagaming-labeling-v2-work	hf → corpus/chunk_rows_to_documents → join → map_column → project	eval only
pt_subset_tokens_human_labeled	camgeodesic/metagaming-labeling-v2-work	stateful_filter → stateful_filter → project	one per doc_label value
regex_dense_docs_human_labeled	camgeodesic/metagaming-labeling-v2-work	pretraining_subset/documents_by_window → project → project → hf → hf → concat → semijoin → stateful_filter → stateful_filter → corpus/chunk_label_lists → corpus/chunk_rows_to_documents → join → map_column → project → corpus/human_labelled_subset → map_column → project	one per split value
regex_dense_docs_opus_judged	camgeodesic/metagaming-labeling-v2-work	pretraining_subset/candidate_stack → stateful_filter → project → project → json_file → hf → semijoin → semijoin → corpus/chunk_rows_to_documents → join → map_column → project	eval only
regex_dense_tokens_human_labeled	camgeodesic/metagaming-labeling-v2-work	stateful_filter → stateful_filter → project	one per doc_label value

Provenance

ctl_dolci_instruct

Source: camgeodesic/metagaming-labeling

python -m dataset_builder configs/ctl_dolci_instruct.yaml --push

ctl_dolci_think

Source: camgeodesic/metagaming-labeling

python -m dataset_builder configs/ctl_dolci_think.yaml --push

ctl_fineweb

Source: camgeodesic/metagaming-labeling

python -m dataset_builder configs/ctl_fineweb.yaml --push

ctl_swe

Source: camgeodesic/metagaming-labeling

python -m dataset_builder configs/ctl_swe.yaml --push

ctl_swe_tests

Source: camgeodesic/metagaming-labeling

python -m dataset_builder configs/ctl_swe_tests.yaml --push

ctl_wildchat

Source: camgeodesic/metagaming-labeling

python -m dataset_builder configs/ctl_wildchat.yaml --push

low_quality_synthetic_annotated

Source: camgeodesic/metagaming-labeling

python -m dataset_builder configs/low_quality_synthetic_annotated.yaml --push

positive_annotations

Source: camgeodesic/metagaming-labeling

python -m dataset_builder configs/positive_annotations.yaml --push

positive_kimi_agentic_misalignment_vea

Source: camgeodesic/metagaming-labeling-v2 Transform: map_column → flat_map → flat_map → project → llm_render_column → aggregate → filter → project

python -m dataset_builder configs/positive_kimi_agentic_misalignment_vea.yaml --push

positive_kimi_metagame_vea

Source: camgeodesic/metagaming-labeling-v2 Transform: map_column → flat_map → flat_map → project → llm_render_column → aggregate → filter → project

python -m dataset_builder configs/positive_kimi_metagame_vea.yaml --push

positive_needham_eval_grading

Source: camgeodesic/metagaming-labeling

python -m dataset_builder configs/positive_needham_eval_grading.yaml --push

positive_westover_labels

Source: camgeodesic/metagaming-labeling-v2 Transform: project → flat_map → hf → semijoin → project → join → project → join → flat_map → hf → semijoin → project → join → project → join → flat_map → hf → semijoin → project → join → project → join → hf → project → join → filter → map_column → filter → filter → project → join → filter → map_column → filter → filter → project → join → map_column → hook → map_column → project

python -m dataset_builder configs/positive_westover_labels.yaml --push

pt_subset_docs_jev_judged

Source: camgeodesic/metagaming-labeling-v2-work Transform: hf → corpus/chunk_rows_to_documents → join → map_column → project

python -m dataset_builder configs/pt_subset_docs_jev_judged.yaml --push

pt_subset_tokens_human_labeled

Source: camgeodesic/metagaming-labeling-v2-work Transform: stateful_filter → stateful_filter → project

python -m dataset_builder configs/pt_subset_tokens_human_labeled.yaml --push

regex_dense_docs_human_labeled

Source: camgeodesic/metagaming-labeling-v2-work Transform: pretraining_subset/documents_by_window → project → project → hf → hf → concat → semijoin → stateful_filter → stateful_filter → corpus/chunk_label_lists → corpus/chunk_rows_to_documents → join → map_column → project → corpus/human_labelled_subset → map_column → project

python -m dataset_builder configs/regex_dense_docs_human_labeled.yaml --push

regex_dense_docs_opus_judged

Source: camgeodesic/metagaming-labeling-v2-work Transform: pretraining_subset/candidate_stack → stateful_filter → project → project → json_file → hf → semijoin → semijoin → corpus/chunk_rows_to_documents → join → map_column → project

python -m dataset_builder configs/regex_dense_docs_opus_judged.yaml --push

regex_dense_tokens_human_labeled

Source: camgeodesic/metagaming-labeling-v2-work Transform: stateful_filter → stateful_filter → project

python -m dataset_builder configs/regex_dense_tokens_human_labeled.yaml --push

Reproducibility

All splits use split_hash() (MD5-based, seeded) so rebuilding from the same config against the same source data produces identical partitions. For an LLM-generated dataset, a provider's seed parameter is best-effort; pin consumer loads to a specific HF commit SHA to avoid drift when the builder re-pushes.

---

This card is auto-generated by dataset_builder.cards.