List your company

Harmlessness Evaluation on HarmBench

76.6Win Ratefsd-var
22.62436.63750.6564.663Mar 11, 2026

Evaluation results

MethodLinks
fsd-varBaseline Model=SFT, Judge Model=GPT-4o-mini2026.0376.610.612.7
fsd-spectral-linearBaseline Model=SFT, Judge Model=GPT-4o-mini2026.0375.312.312.4
fsd-spectral-exponentialBaseline Model=SFT, Judge Model=GPT-4o-mini2026.0374.712.612.7
fsd-cvarBaseline Model=SFT, Judge Model=GPT-4o-mini2026.0372.915.711.4
fsd-spectral-powerBaseline Model=SFT, Judge Model=GPT-4o-mini2026.037114.814.1
Safe RLHFBaseline Model=SFT, Judge Model=GPT-4o-mini2026.0368.120.411.5
fsd-spectral-wangBaseline Model=SFT, Judge Model=GPT-4o-mini2026.0367.313.419.3
fsd-uniformBaseline Model=SFT, Judge Model=GPT-4o-mini2026.0367.17.825.1
fsd-varBaseline Model=Safe RLHF, Judge Model=GPT-4o-mini2026.0339.628.631.8
fsd-uniformBaseline Model=Safe RLHF, Judge Model=GPT-4o-mini2026.033822.239.8
fsd-spectral-linearBaseline Model=Safe RLHF, Judge Model=GPT-4o-mini2026.0334.23728.9
fsd-spectral-powerBaseline Model=Safe RLHF, Judge Model=GPT-4o-mini2026.0330.448.920.7
fsd-cvarBaseline Model=Safe RLHF, Judge Model=GPT-4o-mini2026.0328.646.524.9
fsd-spectral-exponentialBaseline Model=Safe RLHF, Judge Model=GPT-4o-mini2026.032852.219.9
fsd-spectral-wangBaseline Model=Safe RLHF, Judge Model=GPT-4o-mini2026.0324.749.725.5