evaluation-package · Source-linked discovery

OR-Bench

OR-Bench Hard-1K measures over-refusal in LLMs: the tendency to reject safe prompts that superficially resemble harmful ones. The 1K subset contains prompts rejected by at least 3 frontier models, making it the most challenging split of the 80K dataset. Prompts were generated by rewriting toxic seeds into benign borderline variants using Mixtral, then filtered via an ensemble LLM moderator. Scoring uses an LLM judge classifying responses as direct_answer, direct_refusal, or indirect_refusal; the metric is rejection rate (lower is better).