Safety
AIR-Bench 2024
AIR-Bench tests model replies against a taxonomy of risks drawn from government regulations and company policies. This import preserves the published refusal/safety grades, including intermediate values.
7,723items
66subjects
unknownlicense
safetydomain
textmodality
item-level responses released
Saturation status: Yes
Response matrix
Responses come from Stanford CRFM's HELM AIR-Bench leaderboard, which is the benchmark's own evaluation venue rather than a third-party rerun; the panel has grown beyond the 22 models reported by Zeng et al.
Loading response matrix…
lowhighUnobserved
Scale: Score in [0, 1]