Datasetpublic

dougalldeepmind/2026-08-25-table2-9284-difficult-advice-verbose-token-matched-train-mixture

Token-matched verbose difficult-advice arm. Holds difficult advice's share of the TRAINABLE TOKENS at the control's value while the traces are ~3x longer, by keeping only a subset of the expanded rows. Its sibling arm holds the ROW share instead; together they separate more deliberation from more difficult-advice signal. field value experiment Token-matched verbose difficult-advice arm. Holds difficult advice's share of the TRAINABLE TOKENS at the control's value… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-08-25-table2-9284-difficult-advice-verbose-token-matched-train-mixture.

sourceHugging Faceupdated 1mo agoView on Hugging Face

0likes142downloads

Dataset Card

Token-matched verbose difficult-advice arm. Holds difficult advice's share of the TRAINABLE TOKENS at the control's value while the traces are ~3x longer, by keeping only a subset of the expanded rows. Its sibling arm holds the ROW share instead; together they separate more deliberation from more difficult-advice signal.

field	value
`experiment`	Token-matched verbose difficult-advice arm. Holds difficult advice's share of the TRAINABLE TOKENS at the control's value while the traces are ~3x longer, by keeping only a subset of the expanded rows. Its sibling arm holds the ROW share instead; together they separate more deliberation from more difficult-advice signal.
`date_generated`	2026-08-25
`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); pinned to first-party endpoints via configs/endpoints/providers.yaml
`generation_config`	scratch/verbosecot/buildtokenmatchedmixture.py, seed 0. Subset drawn trait-stratified and randomly within trait, from the EXPANDED rows only, accumulating in a trait round-robin until the assistant-token total is closest to the control's.
`schema`	t29284daverbosetokenmatched.jsonl: text (rendered Qwen chat), source, scenarioid, traitid - identical schema to LASR-Callum/2026-08-14-table2-9284-difficult-advice-716-train.
`provenance`	uv run python scratch/verbosecot/buildtokenmatchedmixture.py --push
`composition`	9,647 rows = 363 difficult-advice + 9,284 table2. Difficult-advice trainable tokens 833,388 against the control's 832,780 (1.0007x), i.e. 28.65% of trainable tokens against the control's 28.63%. Row share falls to 3.76% from 7.16% - that is the variable this arm lets float, deliberately. Per trait: t1 41, t2 41, t3 41, t4 40, t5 40, t6 40, t7 40, t8 40, t9 40.
`control_arm`	LASR-Callum/2026-08-14-table2-9284-difficult-advice-716-train - same table2 rows, same difficult-advice token budget, original short traces.
`sibling_arm`	LASR-Callum/2026-08-25-table2-9284-difficult-advice-verbose-716-train - all 716 verbose rows, row share held at 7.16%, token share allowed to rise.