MedSafetyBench

MedSafetyBench evaluates whether a medical LLM honours the principles of medical ethics. It contains harmful medical requests (fabricating records, prescribing without indication, breaching confidentiality, and so on) mapped to the AMA's principles, with paired safe responses for fine tuning. It's from the AI4LIFE group at Harvard.

Narrow by design: useful if you deploy a model in a health context.