Safety Evaluation on BeaverTails (test)
98Blue Reward (Higher)fsd-spectral-exponential
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| fsd-spectral-exponentialCompetitor (Red)=safe-rlhf2026.03 | 98 | 94 | 96 | 93 | |
| fsd-cvarCompetitor (Red)=sft2026.03 | 98 | 92 | 97 | 55 | |
| fsd-spectral-exponentialCompetitor (Red)=sft2026.03 | 98 | 66 | 96 | 55 | |
| fsd-spectral-linearCompetitor (Red)=safe-rlhf2026.03 | 97 | 93 | 94 | 93 | |
| fsd-spectral-wangCompetitor (Red)=safe-rlhf2026.03 | 97 | 95 | 96 | 93 | |
| fsd-spectral-powerCompetitor (Red)=safe-rlhf2026.03 | 97 | 94 | 96 | 93 | |
| fsd-spectral-linearCompetitor (Red)=sft2026.03 | 97 | 74 | 94 | 55 | |
| fsd-spectral-wangCompetitor (Red)=sft2026.03 | 97 | 75 | 96 | 55 | |
| fsd-spectral-powerCompetitor (Red)=sft2026.03 | 97 | 76 | 95 | 55 | |
| fsd-uniformCompetitor (Red)=safe-rlhf2026.03 | 96 | 83 | 92 | 93 | |
| fsd-cvarCompetitor (Red)=safe-rlhf2026.03 | 96 | 98 | 97 | 93 | |
| safe-rlhfCompetitor (Red)=sft2026.03 | 96 | 61 | 93 | 55 | |
| fsd-uniformCompetitor (Red)=sft2026.03 | 95 | 65 | 92 | 55 | |
| fsd-varCompetitor (Red)=sft2026.03 | 93 | 82 | 88 | 55 | |
| fsd-varCompetitor (Red)=safe-rlhf2026.03 | 88 | 88 | 88 | 93 |