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.