geodesic-research/control-pretraining-ai-discourse

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("geodesic-research/control-pretraining-ai-discourse", "<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 dataset holds

AI and machine-learning discourse from the control-pretraining corpora, and the documents two safety filters took out of a model's training data. Every config is one train split, and every document carries id, source, text and num_tokens (the Nemotron base tokenizer's count of text, without an end-of-document token).

config	what it holds
climbmix_ai_docs	NVIDIA ClimbMix's documents selected by an AI-discourse regex gate, each with the rule (flag_category) and the text (flag_term) that selected it
zyda_ai_docs	the same gate over Zyda-2's 100B-token sample
nemotron_wiki_rewrite_ai_docs	the same gate over the Wikipedia rewrite of NVIDIA's Nemotron pretraining data
arxiv_papers	arXiv's cs.AI, cs.CV, cs.LG and stat.ML papers as their own LaTeX, bibliography and preamble removed
ai_safety_and_adjacent	LessWrong with the Alignment Forum, the EA Forum, the Stampy and ARD articles, rewrites of LessWrong posts, and AI labs' risk reports
filtered_broad	every document the broad filter took out of the Broadly Filtered model's pretraining and midtraining data
filtered_narrow	every document the end-to-end narrow filter took out of the Narrowly Filtered V2 E2E model's pretraining and midtraining data

The first five are copied unchanged from geodesic-research/control-pretraining-datasets at revision b367be41, so their ids join back to it.

The filters. The broad filter removes a document when a canary check flags it or the GPT-5 Mini judge scores it 2 or more. The narrow filter removes a document when a canary check flags it or the GPT-5.5 judge scores it 4 or more, and the V2 E2E model applied it to both training stages. Each filtered set stacks the filtered side of every corpus its model trained on and keeps one row per distinct text (by SHA-256). A document's provenance lists every filtered split its text was found in, as {stage, corpus, split}, and its id, source and num_tokens are those of its first occurrence. A text filtered from one copy can still have been trained on in another copy the filter kept; those texts are included, so each set is everything its filter took out. filtered_narrow also carries the judge's verdict on each document: its score and reason per category, the categories it flagged and the canary check's result (the category section below).

Composition — documents and Nemotron tokens per config

Computed from the published data on every card regeneration. Totals are column sums per config, and configs share documents: the filtered sets draw on the corpora the other configs come from, so summing across rows overstates unique volume.

subset	documents	total tokens	mean	median	p95	max
ai_safety_and_adjacent	352,949	658,148,626	1,865	1,042	5,478	1,304,786
arxiv_papers	433,714	8,000,009,008	18,445	14,118	42,035	37,014,723
climbmix_ai_docs	13,506,352	15,892,372,146	1,177	659	3,188	988,937
filtered_broad	577,689	3,034,476,419	5,253	963	23,984	2,799,888
filtered_narrow	50,123	338,011,332	6,744	1,101	29,266	1,883,344
nemotron_wiki_rewrite_ai_docs	53,041	188,829,967	3,560	1,549	12,983	982,969
zyda_ai_docs	1,536,755	4,550,102,536	2,961	1,119	11,727	450,565

Narrow filter — judge categories

The GPT-5.5 judge reads each document and scores it from 1 to 5 in each of the 22 categories of its rubric, for how strongly the document warrants filtering for that reason: 1 is clearly not relevant, 2 relevant but probably not filter-worthy, 3 borderline and leaning no, 4 borderline and leaning yes, 5 clearly yes. A category scored 1 is left out of the verdict, so the scores in the data run from 2 to 5. The narrow filter removes a document when the canary check flags it or the judge's overall score is 4 or more. A text found in several training splits was judged in each, and the judge columns keep each score's highest value with the reason given for it.

column	what it holds
judge_flagged_categories	the categories scored 4 or more, the filter's threshold; empty when the canary check alone removed the document, or when no category reached 4
judge_scores	a dict with the 22 category names as keys and the category's score, 2 to 5, as the value; null where the judge did not list the category
judge_reasons	a dict with the same keys and the judge's one-sentence reason for that score as the value; null where the score is
judge_score	the overall score the filter tested: 4 or 5 for a document the judge removed (the source's record, not always the highest category score), null for one the canary check alone removed
canary	whether the canary check flagged the document

Below, each category's documents and Nemotron tokens, recomputed from the published data on every card regeneration. A document can count in several categories, or in none. The descriptions summarise the rubric's category definitions; the prompt the judge ran is not stored in this repository, and its sha256 (aa59f1f0…) is listed on the card of sudoers/control-pretraining-filter-annotated, so the rubric's wording may differ.

category	what it covers	documents	total tokens
sabotage_strategy	Strategies for an AI to sabotage its developers' research, evaluations or training from within	17,539	130,175,161
model_card	Official model cards, system cards, constitutions, RSPs and safety frameworks, and substantive reviews of them	10,552	15,673,168
training_sabotage	Deliberately manipulating one's own training, or poisoning the data of future models	8,857	108,529,780
monitor_evasion	Concrete ways an AI could evade its developers' safety monitors (CoT monitors, probes)	8,558	47,173,292
dangerous_capability_evals	Benchmarks and system-card evaluations of scheming- or sabotage-relevant actions	8,268	32,636,129
monitoring	Safety monitors that inspect an agent's reasoning or actions: CoT monitoring, probes, action blocking	6,945	39,339,609
scheming_evals	Evaluations of deliberate deception, sandbagging or collusion ability	6,790	26,181,537
jailbreak	Concrete working jailbreak prompts or payloads that bypass safety guardrails	4,906	55,912,244
control_evals	Results on how well AI-control techniques work in practice	4,697	27,174,554
data_poisoning	Deliberate data poisoning that induces a chosen hidden or harmful behaviour	4,368	71,946,665
situational_awareness	Evaluations or demonstrations of situational awareness, introspection or evaluation-awareness	3,341	14,831,187
prod_control	Control measures used or empirically evaluated for deployed frontier AI agents	3,109	15,968,743
agent_cyber	Evaluations of how well AI agents do offensive cyber tasks (autonomous hacking, CTFs)	2,963	12,453,000
other	Other information useful to a misaligned AI for subverting oversight	2,948	51,224,179
rogue_strategy	Actionable strategies for an AI to escape its control regime: self-exfiltration, rogue deployment	2,801	14,619,449
collusion	AI-to-AI collusion against oversight: steganography, secret signals, Schelling points	2,460	21,629,616
lab_security	Frontier-lab cybersecurity: egress limits, incident response, weight security	1,455	5,081,424
honeypots	Honeypots for detecting malicious AI activity, and how to build them	660	4,992,274
false_facts_sdf	Instilling false beliefs, including synthetic-document finetuning	620	6,684,127
misaligned_transcripts	First-person transcripts of models pursuing or describing strategic misalignment	591	7,469,015
gradient_hacking	Results on the efficacy of gradient hacking	36	183,324
exploration_hacking	Results on the efficacy of exploration hacking	35	146,932

Configs

Config	Source	Transform	Splits
ai_safety_and_adjacent	geodesic-research/control-pretraining-datasets	none	none
arxiv_papers	geodesic-research/control-pretraining-datasets	none	none
climbmix_ai_docs	geodesic-research/control-pretraining-datasets	none	none
filtered_broad	geodesic-research/control-pretraining-datasets	hf/stack_subsets → map_column → aggregate → project	none
filtered_narrow	geodesic-research/control-pretraining-datasets	hf/stack_subsets → hf/stack_subsets → map_column → map_column → aggregate → map_column → map_column → map_column → project	none
nemotron_wiki_rewrite_ai_docs	geodesic-research/control-pretraining-datasets	none	none
zyda_ai_docs	geodesic-research/control-pretraining-datasets	none	none

Provenance

ai_safety_and_adjacent

Source: geodesic-research/control-pretraining-datasets

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

arxiv_papers

Source: geodesic-research/control-pretraining-datasets

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

climbmix_ai_docs

Source: geodesic-research/control-pretraining-datasets

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

filtered_broad

Source: geodesic-research/control-pretraining-datasets Transform: hf/stack_subsets → map_column → aggregate → project

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

filtered_narrow

Source: geodesic-research/control-pretraining-datasets Transform: hf/stack_subsets → hf/stack_subsets → map_column → map_column → aggregate → map_column → map_column → map_column → project

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

nemotron_wiki_rewrite_ai_docs

Source: geodesic-research/control-pretraining-datasets

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

zyda_ai_docs

Source: geodesic-research/control-pretraining-datasets

python -m dataset_builder configs/zyda_ai_docs.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.publishing.cards.