Views
No views yet
Difficult-advice SFT corpus, GPT responder-swap arm: the baseline's own scenarios and prompts held frozen and ONLY the assistant's reply regenerated by OpenAI models. Stages 1-4 (constitution chunking, scenarios, prompt drafting, prompt revision) are the published baseline run's, reused verbatim; openai/gpt-5.6-luna writes the draft response and openai/gpt-5.6-terra revises it against the full constitution. Third arm of the generator ablation, alongside the Anthropic baseline and… See the full description on the dataset page: https://huggingface.co/datasets/LASR-Callum/2026-08-25-difficult-advice-gpt-responder-716.