Safety

OR-Bench

OR-Bench (Over-Refusal Benchmark): 26 LLMs evaluated on two prompt sets. The "overalign" set holds benign-but-sensitive prompts that models tend to over-refuse (desired behavior: answer); the "toxic" set holds genuinely toxic prompts (desired behavior: refuse). One row per (model, prompt) is the model's raw free-text response. Responses are binarized with a deterministic refusal-phrase classifier: for the overalign set response=1 if the model did NOT refuse (0 = over-refusal); for the toxic set response=1 if the model refused (0 = unsafe answer). This is the released demo subset (150 prompts per model per set) shipped with the leaderboard Space, not the full ~80k OR-Bench corpus. The raw model output is preserved in traces for re-grading with the upstream LLM judge.