Datasetpublic

dougalldeepmind/2026-08-04-table2-instruction-tuning-mixture-spec-filtered

Table 2 instruction-tuning mixture, spec-filtered A reproduction of the paper's Table 2 instruction-tuning mixture at its exact per-source sample counts, plus an LLM spec-alignment filter and the per-sample judge verdicts, so the filter can be re-cut at any threshold without paying to re-judge. field value experiment Table 2 instruction-tuning mixture for the Teaching Claude Why replication, filtered for spec misalignment date_generated 2026-08-04 constitution… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-08-04-table2-instruction-tuning-mixture-spec-filtered.

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes100downloads

Dataset Card

Table 2 instruction-tuning mixture, spec-filtered

A reproduction of the paper's Table 2 instruction-tuning mixture at its exact per-source sample counts, plus an LLM spec-alignment filter and the per-sample judge verdicts, so the filter can be re-cut at any threshold without paying to re-judge.

field	value
`experiment`	Table 2 instruction-tuning mixture for the Teaching Claude Why replication, filtered for spec misalignment
`date_generated`	2026-08-04
`constitution`	`claude_distilled_12_principles_mid` — 9 numbered principles (the "12" is historical). The full text was given to the judge on every call.
`source_repo`	`teaching_claude_why_replication` @ `96ff8aa36dd7d9914b016f88e80b1e3724f96ccc`
`models`	judge: `openai/gpt-5.6-terra` via OpenRouter
`generation_config`	temperature 0.0, `reasoning_effort` low, max_tokens 900, seed 0 for sampling
`schema`	see below
`provenance`	`scratch/build_paper_mixture.py --config configs/data/mixture_paper_table2.yaml` then `scratch/filter_spec_misaligned.py --mixture <mixture.jsonl>`

Files

file	rows	what
`mixture_unfiltered.jsonl`	10,000	the mixture as built, before filtering
`mixture_filtered.jsonl`	9,285	after the spec filter — the training file
`verdicts.jsonl`	10,000	one verdict per sample: `idx`, `source`, `verdict`, `category`, `why`, `parsed`
`filter_report.json`	—	aggregate counts by source and category
`mixture_stats.json`	—	per-source token counts of the unfiltered build

mixture_*.jsonl: text (Qwen3.6 chat-template-rendered, no think block) and source.

Composition (unfiltered, Table 2 counts verbatim)

No Robots 2,779 · Tulu3 IF 1,471 · Self-Oss-Instruct 1,064 · NuminaMath CoT 1,063 · Smol-constraints 1,055 · APIGen-Function-Calling 1,054 · Smol-summarize 984 · LIMA 314 · LongAlign 216 — 10,000 samples, 5,423,299 Qwen3.6 tokens.

Filtering: 715 of 10,000 rejected (7.15%)

Rejected by category: {'spec_violation': 683, 'toxic': 30, 'no_self_disclaimer': 2}

source	total	kept	reject %
smol_summarize	984	669	32.01%
tulu3_if	1,471	1,365	7.21%
lima	314	292	7.01%
apigenfunctioncalling	1,054	987	6.36%
no_robots	2,779	2,640	5.0%
numinamath_cot	1,063	1,037	2.45%
smol_constraints	1,055	1,034	1.99%
selfossinstruct	1,064	1,047	1.6%
longalign	216	214	0.93%

Read this before using it as a "spec filter"

The paper describes the step as catching "toxic data, data where AI identifies itself as another model (e.g. 'I'm GPT-4') or claim 'As an AI, I have no subjective opinions/preferences'." This filter found zero of the latter two. Rejections are almost entirely spec_violation, and roughly half of those are instruction-following failures — a summary ignoring its own three-sentence limit, or using pronouns the prompt forbade. That is a real data-quality problem, but it is a broader screen than the paper describes, and it falls unevenly: smol_summarize loses 31% while longalign loses 0.9%.

Because verdicts.jsonl carries a category per sample, a paper-faithful cut (toxic / wrong_identity / no_self_disclaimer only, ~0.3% removed) can be rebuilt from these files with no further judging.

79 judge replies did not parse and were kept, flagged with parsed: false. Filtering on unparseable output would silently delete good data.

Caveats

—LongAlign is not representative: its rows average 10,677 tokens and only the shortest ~18% fit under the 8,192 cap, so 965 were dropped to find 216. Those 216 are 2.2% of samples but 29% of tokens, and only ~2% of their tokens are supervised.
—LIMA came from `64bits/lima_vicuna_format`, a third-party re-upload, because GAIR/lima ships a loading script the Hub no longer supports.
—Rendered with no think block. The marked variant used for training (empty <think></think> on non-reasoning rows, masked out of the loss) is in `2026-08-04-table2-synthdoc-h200x4-train`.
—Not evaluated.