Datasetpublic

dougalldeepmind/2026-08-27-table2-9284-good-ai-fiction-716-train

Table2 9,284 + Good AI Fiction 716 — SFT training mixture field value experiment The fiction arm of the alignment-data comparison: the SAME 9,284 benign capability-preserving rows the difficult-advice mixture uses, with its 716 difficult-advice rows replaced by 716 first-person Good AI Fiction rows at a matched trainable-token budget. Train against LASR-Callum/2026-08-14-table2-9284-difficult-advice-716-train to read the difference as content, not size.… See the full description on the dataset page: https://huggingface.co/datasets/dougalldeepmind/2026-08-27-table2-9284-good-ai-fiction-716-train.

sourceHugging Faceupdated 1mo agoView on Hugging Face

0likes123downloads

Dataset Card

Table2 9,284 + Good AI Fiction 716 — SFT training mixture

field	value
`experiment`	The fiction arm of the alignment-data comparison: the SAME 9,284 benign capability-preserving rows the difficult-advice mixture uses, with its 716 difficult-advice rows replaced by 716 first-person Good AI Fiction rows at a matched trainable-token budget. Train against LASR-Callum/2026-08-14-table2-9284-difficult-advice-716-train to read the difference as content, not size.
`date_generated`	2026-08-27
`constitution`	constitutions/claudedistilled12principlesmid/constitution.md (the fiction half only; the benign half has none)
`source_repo`	teachingclaudewhy_replication @ ae0725130a2fccd74fe7bdef5c570ec71420cd7b
`models`	anthropic/claude-haiku-4.5 (scenarios, prompts); anthropic/claude-sonnet-5 (story, rewrite); openai/gpt-5.6-terra (constitution-aware critic); x-ai/grok-4.6 (two independent accept gates, pattern scan); benign half regenerated by nobody — replayed verbatim
`generation_config`	temperature 1.1 scenarios / 1.0 prompts / 0.9 story / 0.8 rewrite / 0.0 judges; max_tokens 8192 scenarios, 2048 prompts, 8192 story, 12288 rewrite; seed 0; providers pinned per configs/endpoints/providers.yaml
`schema`	JSONL, pre-rendered Qwen chat form. `text`: the full conversation as `<	im_start	>{role}\n{content}<	im_end	>\n `per turn, the assistant turn carrying` <think>\n{reasoning}\n</think>\n\n{answer} `(an EMPTY marker on the benign rows, a real trace on the fiction rows).` source`:` goodaifiction `or the benign source name. Fiction rows also carry` traitid` and `scenarioid`.
`provenance`	uv run python scratch/goodaifiction/publish.py mixture --run <run dir>, which calls scratch/buildt29284da716mixture.py with --synthrepo LASR-Callum/2026-08-27-good-ai-fiction-716 and --idsfrom the audited 716-row selection. Benign half pinned to the difficult-advice arm's own source, unmodified.
`rows`	10000
`composition`	716 goodaifiction + 9284 benign
`synth_fraction`	0.0716
`alignment_trainable_tokens`	822,424 over 716 rows (difficult-advice slice: 832,064)
`benign_half_source`	LASR-Callum/2026-08-04-table2-instruction-tuning-9284-filtered-8192::mixture_think.jsonl — unchanged
`paired_arm`	LASR-Callum/2026-08-14-table2-9284-difficult-advice-716-train
`alignment_subset`	LASR-Callum/2026-08-27-good-ai-fiction-716