Datasets:

sudoers
/

control-pretraining-filter-annotated

---
configs:
- config_name: climbmix_full
  data_files:
  - split: train
    path: climbmix_full/*.parquet
- config_name: climbmix_ai_docs
  data_files:
  - split: train
    path: climbmix_ai_docs/*.parquet
- config_name: climbmix_long
  data_files:
  - split: train
    path: climbmix_long/*.parquet
- config_name: climbmix_ai_docs_long
  data_files:
  - split: train
    path: climbmix_ai_docs_long/*.parquet
- config_name: zyda_full
  data_files:
  - split: train
    path: zyda_full/*.parquet
- config_name: zyda_ai_docs
  data_files:
  - split: train
    path: zyda_ai_docs/*.parquet
- config_name: zyda_long
  data_files:
  - split: train
    path: zyda_long/*.parquet
- config_name: zyda_ai_docs_long
  data_files:
  - split: train
    path: zyda_ai_docs_long/*.parquet
- config_name: ai_safety_and_adjacent
  data_files:
  - split: train
    path: ai_safety_and_adjacent/*.parquet
- config_name: stack_edu
  data_files:
  - split: train
    path: stack_edu/*.parquet
- config_name: stack_edu_long
  data_files:
  - split: train
    path: stack_edu_long/*.parquet
- config_name: nemotron_wiki_rewrite
  data_files:
  - split: train
    path: nemotron_wiki_rewrite/*.parquet
- config_name: nemotron_wiki_rewrite_ai_docs
  data_files:
  - split: train
    path: nemotron_wiki_rewrite_ai_docs/*.parquet
- config_name: arxiv_papers
  data_files:
  - split: train
    path: arxiv_papers/*.parquet
- config_name: nemotron_stem_sft
  data_files:
  - split: train
    path: nemotron_stem_sft/*.parquet
- config_name: pa_warm_start_sft
  data_files:
  - split: train
    path: pa_warm_start_sft/*.parquet
- config_name: pa_warm_start_xl
  data_files:
  - split: train
    path: pa_warm_start_xl/*.parquet
---
# control-pretraining-filter-annotated

- `climbmix_full` — from [geodesic-research/control-pretraining-datasets](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/geodesic-research/control-pretraining-datasets)
- `stack_edu` — from [geodesic-research/control-pretraining-datasets](https://huggingface.co/datasets/geodesic-research/control-pretraining-datasets)
- `stack_edu_long` — from [geodesic-research/control-pretraining-datasets](https://huggingface.co/datasets/geodesic-research/control-pretraining-datasets)
- `nemotron_wiki_rewrite` — from [geodesic-research/control-pretraining-datasets](https://huggingface.co/datasets/geodesic-research/control-pretraining-datasets)
- `nemotron_wiki_rewrite_ai_docs` — from [geodesic-research/control-pretraining-datasets](https://huggingface.co/datasets/geodesic-research/control-pretraining-datasets)
- `arxiv_papers` — from [geodesic-research/control-pretraining-datasets](https://huggingface.co/datasets/geodesic-research/control-pretraining-datasets)
- `nemotron_stem_sft` — from [geodesic-research/control-pretraining-datasets](https://huggingface.co/datasets/geodesic-research/control-pretraining-datasets)
- `pa_warm_start_sft` — from [geodesic-research/pa-warm-start-sft-heavy-25b-mix](https://huggingface.co/datasets/geodesic-research/pa-warm-start-sft-heavy-25b-mix)
- `pa_warm_start_xl` — from [geodesic-research/pa-warm-start-sft-xl-50b-mix](https://huggingface.co/datasets/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

Rates vary by two to three orders of magnitude across configs, so there is no single
headline number — compute per config. Document-level `filter_decision` rates, measured
2026-09-22 after every config was fully judged:

| config | documents | production (judge>=4) | aggressive (mini>=4) |
|---|---:|---:|---:|
| `arxiv_papers` | 433,714 | 2.4203% | 9.868% |
| `climbmix_ai_docs_long` | 5,801 | 0.7068% | 3.7924% |
| `zyda_ai_docs_long` | 1,665 | 0.1802% | 2.6426% |
| `climbmix_long` | 800,032 | 0.0694% | 0.5351% |
| `climbmix_ai_docs` | 13,506,352 | 0.0433% | 0.4834% |
| `zyda_long` | 139,223 | 0.0345% | 0.4726% |
| `stack_edu_long` | 3,190 | 0.0313% | 0.5956% |
| `pa_warm_start_sft` | 5,702,903 | 0.0143% | 0.4606% |
| `pa_warm_start_xl` | 8,924,246 | 0.0075% | 0.3800% |
| `climbmix_full` | 553,315,056 | 0.0011% | 0.0119% |

Filtered documents run longer than the corpus mean, so token-level rates exceed
document-level rates (e.g. `pa_warm_start_xl`: 0.0075% of docs, 0.0076% of tokens
for production; 0.38% of docs but 0.9485% of tokens for aggressive).

### 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` — none remain
Earlier revisions of this dataset contained rows marked `decided_by: "unscored"`: they passed
the CPU prefilter and reached `mini_score >= 4`, but the run stopped at its spend ceiling before
the judge saw them, so their `filter_decision = false` meant **"not established"** rather than
"judged and kept". **As of 2026-09-22 every such row has been judged and no config contains any
unscored rows.** `filter_decision = false` now uniformly means the cascade completed and did not
flag the document. If you cached a copy before that date, re-pull it — `arxiv_papers` was the
most affected (9.7% of its rows), and its production filter rate changed from ~0% to 2.42%.

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.

	---
	configs:
	- config_name: climbmix_full
	data_files:
	- split: train
	path: climbmix_full/*.parquet
	- config_name: climbmix_ai_docs
	data_files:
	- split: train
	path: climbmix_ai_docs/*.parquet
	- config_name: climbmix_long
	data_files:
	- split: train
	path: climbmix_long/*.parquet
	- config_name: climbmix_ai_docs_long
	data_files:
	- split: train
	path: climbmix_ai_docs_long/*.parquet
	- config_name: zyda_full
	data_files:
	- split: train
	path: zyda_full/*.parquet
	- config_name: zyda_ai_docs
	data_files:
	- split: train
	path: zyda_ai_docs/*.parquet
	- config_name: zyda_long
	data_files:
	- split: train
	path: zyda_long/*.parquet
	- config_name: zyda_ai_docs_long
	data_files:
	- split: train
	path: zyda_ai_docs_long/*.parquet
	- config_name: ai_safety_and_adjacent
	data_files:
	- split: train
	path: ai_safety_and_adjacent/*.parquet
	- config_name: stack_edu
	data_files:
	- split: train
	path: stack_edu/*.parquet
	- config_name: stack_edu_long
	data_files:
	- split: train
	path: stack_edu_long/*.parquet
	- config_name: nemotron_wiki_rewrite
	data_files:
	- split: train
	path: nemotron_wiki_rewrite/*.parquet
	- config_name: nemotron_wiki_rewrite_ai_docs
	data_files:
	- split: train
	path: nemotron_wiki_rewrite_ai_docs/*.parquet
	- config_name: arxiv_papers
	data_files:
	- split: train
	path: arxiv_papers/*.parquet
	- config_name: nemotron_stem_sft
	data_files:
	- split: train
	path: nemotron_stem_sft/*.parquet
	- config_name: pa_warm_start_sft
	data_files:
	- split: train
	path: pa_warm_start_sft/*.parquet
	- config_name: pa_warm_start_xl
	data_files:
	- split: train
	path: pa_warm_start_xl/*.parquet
	---

	# control-pretraining-filter-annotated

	- `climbmix_full` — from [geodesic-research/control-pretraining-datasets](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/geodesic-research/control-pretraining-datasets)
	- `stack_edu` — from [geodesic-research/control-pretraining-datasets](https://huggingface.co/datasets/geodesic-research/control-pretraining-datasets)
	- `stack_edu_long` — from [geodesic-research/control-pretraining-datasets](https://huggingface.co/datasets/geodesic-research/control-pretraining-datasets)
	- `nemotron_wiki_rewrite` — from [geodesic-research/control-pretraining-datasets](https://huggingface.co/datasets/geodesic-research/control-pretraining-datasets)
	- `nemotron_wiki_rewrite_ai_docs` — from [geodesic-research/control-pretraining-datasets](https://huggingface.co/datasets/geodesic-research/control-pretraining-datasets)
	- `arxiv_papers` — from [geodesic-research/control-pretraining-datasets](https://huggingface.co/datasets/geodesic-research/control-pretraining-datasets)
	- `nemotron_stem_sft` — from [geodesic-research/control-pretraining-datasets](https://huggingface.co/datasets/geodesic-research/control-pretraining-datasets)
	- `pa_warm_start_sft` — from [geodesic-research/pa-warm-start-sft-heavy-25b-mix](https://huggingface.co/datasets/geodesic-research/pa-warm-start-sft-heavy-25b-mix)
	- `pa_warm_start_xl` — from [geodesic-research/pa-warm-start-sft-xl-50b-mix](https://huggingface.co/datasets/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

	Rates vary by two to three orders of magnitude across configs, so there is no single
	headline number — compute per config. Document-level `filter_decision` rates, measured
	2026-09-22 after every config was fully judged:

	\| config \| documents \| production (judge>=4) \| aggressive (mini>=4) \|
	\|---\|---:\|---:\|---:\|
	\| `arxiv_papers` \| 433,714 \| 2.4203% \| 9.868% \|
	\| `climbmix_ai_docs_long` \| 5,801 \| 0.7068% \| 3.7924% \|
	\| `zyda_ai_docs_long` \| 1,665 \| 0.1802% \| 2.6426% \|
	\| `climbmix_long` \| 800,032 \| 0.0694% \| 0.5351% \|
	\| `climbmix_ai_docs` \| 13,506,352 \| 0.0433% \| 0.4834% \|
	\| `zyda_long` \| 139,223 \| 0.0345% \| 0.4726% \|
	\| `stack_edu_long` \| 3,190 \| 0.0313% \| 0.5956% \|
	\| `pa_warm_start_sft` \| 5,702,903 \| 0.0143% \| 0.4606% \|
	\| `pa_warm_start_xl` \| 8,924,246 \| 0.0075% \| 0.3800% \|
	\| `climbmix_full` \| 553,315,056 \| 0.0011% \| 0.0119% \|

	Filtered documents run longer than the corpus mean, so token-level rates exceed
	document-level rates (e.g. `pa_warm_start_xl`: 0.0075% of docs, 0.0076% of tokens
	for production; 0.38% of docs but 0.9485% of tokens for aggressive).

	### 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` — none remain

	Earlier revisions of this dataset contained rows marked `decided_by: "unscored"`: they passed
	the CPU prefilter and reached `mini_score >= 4`, but the run stopped at its spend ceiling before
	the judge saw them, so their `filter_decision = false` meant "not established" rather than
	"judged and kept". **As of 2026-09-22 every such row has been judged and no config contains any
	unscored rows.** `filter_decision = false` now uniformly means the cascade completed and did not
	flag the document. If you cached a copy before that date, re-pull it — `arxiv_papers` was the
	most affected (9.7% of its rows), and its production filter rate changed from ~0% to 2.42%.

	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.