Safety
HarmBench
HarmBench measures whether a model refuses harmful requests, covering behaviors such as cybercrime, misinformation and chemical or biological harm, with a classifier judging each reply as safe or unsafe. Recorded scores retain intermediate judge ratings; 1 indicates the best rubric score and 0 the worst.
403items
81subjects
unknownlicense
safetydomain
textmodality
item-level responses released
Saturation status: Yes
Response matrix
Responses come from Stanford CRFM's HELM Safety leaderboard (v1.17.0), which prompts each behavior directly and grades with its own safety classifier; the original HarmBench evaluation wraps behaviors in 16 attack methods and judges with its own Llama-2-13B classifier. Complements the standalone harmbench build.
Loading response matrix…
lowhighUnobserved
Scale: Score in [0, 1]