control-pretraining-filter-annotated

• climbmix_full — from geodesic-research/control-pretraining-datasets (stage decisions reused from the run-1 dataset where documents match by content)
• climbmix_ai_docs — from geodesic-research/control-pretraining-datasets (stage decisions reused from the run-1 dataset where documents match by content)
• climbmix_long — from geodesic-research/control-pretraining-datasets (stage decisions reused from the run-1 dataset where documents match by content)
• climbmix_ai_docs_long — from geodesic-research/control-pretraining-datasets (stage decisions reused from the run-1 dataset where documents match by content)
• zyda_full — from geodesic-research/control-pretraining-datasets (stage decisions reused from the run-1 dataset where documents match by content)
• zyda_ai_docs — from geodesic-research/control-pretraining-datasets (stage decisions reused from the run-1 dataset where documents match by content)
• zyda_long — from geodesic-research/control-pretraining-datasets (stage decisions reused from the run-1 dataset where documents match by content)
• zyda_ai_docs_long — from geodesic-research/control-pretraining-datasets (stage decisions reused from the run-1 dataset where documents match by content)
• ai_safety_and_adjacent — from geodesic-research/control-pretraining-datasets
• stack_edu — from geodesic-research/control-pretraining-datasets
• stack_edu_long — from geodesic-research/control-pretraining-datasets
• nemotron_wiki_rewrite — from geodesic-research/control-pretraining-datasets
• nemotron_wiki_rewrite_ai_docs — from geodesic-research/control-pretraining-datasets
• arxiv_papers — from geodesic-research/control-pretraining-datasets
• nemotron_stem_sft — from geodesic-research/control-pretraining-datasets
• pa_warm_start_sft — from geodesic-research/pa-warm-start-sft-heavy-25b-mix
• pa_warm_start_xl — from geodesic-research/pa-warm-start-sft-xl-50b-mix (stage decisions reused from the pa_warm_start_sft config where documents match by content)

Copies of the source datasets listed above (17 configs), with per-stage decisions from a multi-stage "useful to a misaligned AI" pretraining filter added to every row. Original columns are unchanged; each config keeps its source schema. Built 2026-08-10 onward; most recent config added 2026-09-19.

Filter pipeline

canary check -> CPU regex prefilter -> gpt-5-nano relevance (first 1,000,000 chars, stop if < 2) -> gpt-5-mini score (first 1,000,000 chars, stop if < 4) -> gpt-5.5 judge (prompts/judge_new.txt, first 1,000,000 chars). Production decision: filter iff canary or judge >= 4.

Every stage reads the whole document up to that cap; longer documents are truncated to it. Caveats for scores imported by content-hash reuse are in "What the nano gate excluded" below.

Prompt hashes (sha256): nano.txt 8ec4a5f8082da285c5c49e55ec0ef1c53bec4b50d704b54991dfc7beab19234b, mini.txt af2396a96350997533b937c84de19a23e9f925020ccf5489ed82ce1045d748fd, judge_new.txt aa59f1f055f9ec3517014dd7eb13cd5c93ec15f90e7059c260ac34b642394e69.

Added columns

column	type	meaning
prefilter_pass	bool	CPU regex prefilter fired
canary	bool	BigBench canary GUID present (auto-filter; LLM stages skipped)
nano_score	int64 or null	gpt-5-nano score; null = stage did not run
mini_score	int64 or null	gpt-5-mini score; null = stage did not run
judge_score	int64 or null	gpt-5.5 judge_new score; null = stage did not run
judge_categories	JSON string or null	per-category {score, reason} from the judge
decided_by	string	canary / prefilter / nano / mini / judge / unscored
filter_decision	bool	canary or judge_score >= 4 (the conservative/production filter)
filtered_aggressive	bool	canary or mini_score >= 4 — everything the mini tier would have escalated to the judge. A strict superset of filter_decision, and defined even where the judge never ran. Run 2/3 configs only.

What the filter removes

config	documents	filtered docs	doc rate	corpus tokens	filtered tokens	token rate
climbmix	553,240,576	5,886	0.0011%	339.3B	~10.7M	0.0032%
zyda	91,220,256	326	0.0004%	94.8B	~1.1M	0.0012%
lesswrong	67,278	2,087	3.10%	0.40B	20.17M	5.00%

Token counts are o200k_base; lesswrong is exact, the web corpora are estimated from 30 sampled shards each (95% CI: climbmix 4.6-19.8M, zyda 0.1-3.2M). Filtered documents run 1.6-3.3x longer than the corpus mean, so token-level rates exceed document-level rates. Lowering the bar to judge >= 3 would remove ~11.2M additional tokens across the three configs.

What the nano gate excluded (decided_by: nano)

The nano tier is a relevance gate, not a filter: a document scoring < 2 is kept (filter_decision = false) but never escalated to mini or the judge, so it is never examined in depth. Across run 2 that was 5,829,526 of 16,148,647 prefilter survivors (36.1%).

Those documents are identifiable as decided_by == "nano". Treat them as screened out early rather than judged harmless — a nano false negative is invisible downstream. The gate rate tracks content density: 6.4% on ai_safety_and_adjacent versus 38-60% on generic web corpora.

Run 2 gave nano the whole document (up to the model's context window). Judgments imported from run 1 via content-hash reuse were gated by a nano that saw only the first 4,000 characters; the *_long configs were re-scored at full context for that reason, the other reuse configs were not.

decided_by: unscored

Some rows carry decided_by: "unscored": they passed the CPU prefilter but their cascade did not finish, because the run stopped at its spend ceiling. Their filter_decision is false, which means "not established", not "judged and kept" — treat them as unknown rather than as negatives when computing filter rates or training a router. Stage columns show how far they got.

Pick your own threshold downstream: e.g. "filter at judge >= 3" or "route anything with mini_score >= 4". See the source repo's METHOD.md for rubric details. Licenses/terms of the source datasets apply to the copied columns.