abhayesian/

ryan-greenblatt-simulator-eval-questions-v1

25 AI-safety questions designed to evaluate whether a Ryan-Greenblatt-style finetuned simulator answers fresh questions in a way that plausibly represents how Ryan would think and write. Selected from a pool of 368 LLM-generated candidates after embedding-based decontamination against the training corpus and LLM-judge filters for specificity, discrimination, and external-reference / multi-question issues. LOCKED: do not modify… See the full description on the dataset page: https://huggingface.co/datasets/abhayesian/ryan-greenblatt-simulator-eval-questions-v1.