Safety alignment against harmful fine-tuning on BeaverTails
60.9Harm Score (HS)Vaccine
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| VaccineBackbone=LLaMA3-8B2026.05 | 60.9 | 27.4 | |
| BoosterBackbone=LLaMA3-8B2026.05 | 58.1 | 40 | |
| RepNoiseBackbone=LLaMA3-8B2026.05 | 57.2 | 37.9 | |
| Buffer-and-ReinforceBackbone=LLaMA3-8B2026.05 | 8.4 | 77.5 |