Datasetpublic

dougalldeepmind/2026-08-25-table2-9284-difficult-advice-verbose-716-train

Verbose-CoT arm: the published difficult-advice mixture with the 716 difficult-advice reasoning traces expanded ~3x in length, same ideas, to isolate deliberation length from content. field value experiment Verbose-CoT arm: the published difficult-advice mixture with the 716 difficult-advice reasoning traces expanded ~3x in length, same ideas, to isolate deliberation length from content. date_generated 20260825 constitution… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-08-25-table2-9284-difficult-advice-verbose-716-train.

sourceHugging Faceupdated 27d agoView on Hugging Face

0likes55downloads

Dataset Card

Verbose-CoT arm: the published difficult-advice mixture with the 716 difficult-advice reasoning traces expanded ~3x in length, same ideas, to isolate deliberation length from content.

field	value
`experiment`	Verbose-CoT arm: the published difficult-advice mixture with the 716 difficult-advice reasoning traces expanded ~3x in length, same ideas, to isolate deliberation length from content.
`date_generated`	20260825
`constitution`	constitutions/claudedistilled12principlesmid/constitution.md (inherited from the source run; never rendered into any prompt of the expansion itself)
`source_repo`	https://github.com/Matthew-Bozoukov/teachingclaudewhy_replication.git @ d1fa94d14499b20f35215269b5a86ee43fb5eded
`models`	expansion anthropic/claude-sonnet-5 (temp 0.7); fidelity and coverage judges openai/gpt-5.6-terra (temp 0.0); both pinned to first-party endpoints via configs/endpoints/providers.yaml
`generation_config`	configs/data/synth/2026-08-25verbosecot.yaml — ask 4.3x per source paragraph, 170 words per output paragraph, per-record band 2.0-4.5x, 3 attempts then fallback to the original trace
`schema`	t29284da716verbose10k.jsonl: one row per training example, identical to LASR-Callum/2026-08-14-table2-9284-difficult-advice-716-train except that difficultadvicev2 rows carry an expanded <think> block. Fields: text (rendered Qwen chat), source, scenarioid, traitid.
`provenance`	uv run synth run --config configs/data/synth/2026-08-25verbosecot.yaml && uv run python scratch/verbosecot/buildverbose_mixture.py --push
`expansion_outcome`	difficult-advice think words 343,403 -> 962,832 (2.804x overall; the 637 rows that were expanded average 3.03x). 637 expanded, 50 kept their original trace after 3 attempts failed the fidelity or coverage judge, 29 kept it because Anthropic's content filter refused the prompt outright. Every row carries `expansion_status`; the unexpanded ~11% are identical to the control arm and dilute the intervention accordingly.
`composition`	10,000 rows: 716 difficult-advice (7.16% of rows, the SAME row share as the control arm) + 9,284 table2. Holding the row share fixed while the traces got longer moves difficult advice from 32.9% to 48.5% of assistant words and grows total trainable text 1.30x — that shift is a consequence of the design, not a confound that was overlooked. A size-matched control at the ORIGINAL token ratio is a separate arm.
`control_arm`	LASR-Callum/2026-08-14-table2-9284-difficult-advice-716-train — identical but for the difficult-advice think blocks; the 9,284 table2 rows are byte-identical and row order is preserved, verified by verify_substitution() at build time.