Malicious Fine-tuning Defense on BeaverTails (test)
1Harmfulness ScoreDeepAlign
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| DeepAlignBackbone=Llama2-7b-chat, MFT Setting=10-shot2025.07 | 1 | 0 | |
| SimPOBackbone=Llama2-7b-chat, MFT Setting=Initial2025.07 | 1.01 | 0 | |
| VanillaBackbone=Llama2-7b-chat, MFT Setting=Initial2025.07 | 1.06 | 0.3 | |
| SimPOBackbone=Llama2-7b-chat, MFT Setting=10-shot2025.07 | 1.14 | 3.9 | |
| BoosterBackbone=Llama2-7b, MFT Setting=Initial2025.07 | 1.26 | 0 | |
| TARBackbone=Llama2-7b-chat, MFT Setting=Initial2025.07 | 1.26 | 3.3 | |
| DeepAlignBackbone=Llama2-7b-chat, MFT Setting=Initial2025.07 | 1.57 | 0 | |
| SDDBackbone=Llama2-7b-chat, MFT Setting=100-shot2025.07 | 1.57 | 0 | |
| SimPOBackbone=Llama2-7b-chat, MFT Setting=50-shot2025.07 | 1.77 | 12.7 | |
| DeepAlignBackbone=Llama2-7b-chat, MFT Setting=50-shot2025.07 | 2.14 | 14.2 | |
| SDDBackbone=Llama2-7b-chat, MFT Setting=Initial2025.07 | 2.14 | 0 | |
| SDDBackbone=Llama2-7b-chat, MFT Setting=10-shot2025.07 | 2.14 | 0 | |
| SDDBackbone=Llama2-7b, MFT Setting=Initial2025.07 | 2.39 | 15.1 | |
| T-VaccineBackbone=Llama2-7b, MFT Setting=Initial2025.07 | 2.43 | 16.6 | |
| SDDBackbone=Llama2-7b-chat, MFT Setting=50-shot2025.07 | 2.57 | 0 | |
| SDDBackbone=Llama2-7b, MFT Setting=10-shot2025.07 | 2.66 | 15.1 | |
| SDDBackbone=Llama2-7b, MFT Setting=50-shot2025.07 | 2.67 | 21.2 | |
| DeepAlignBackbone=Llama2-7b, MFT Setting=Initial2025.07 | 2.88 | 18.1 | |
| SDDBackbone=Llama2-7b, MFT Setting=100-shot2025.07 | 2.97 | 18.1 | |
| SimPOBackbone=Llama2-7b-chat, MFT Setting=100-shot2025.07 | 3.02 | 34.2 | |
| VanillaBackbone=Llama2-7b, MFT Setting=Initial2025.07 | 3.17 | 26.7 | |
| T-VaccineBackbone=Llama2-7b, MFT Setting=10-shot2025.07 | 3.23 | 33.3 | |
| T-VaccineBackbone=Llama2-7b, MFT Setting=100-shot2025.07 | 3.23 | 26.7 | |
| VanillaBackbone=Llama2-7b, MFT Setting=100-shot2025.07 | 3.38 | 43.3 | |
| DeepAlignBackbone=Llama2-7b-chat, MFT Setting=100-shot2025.07 | 3.43 | 42.7 | |
| SimPOBackbone=Llama2-7b, MFT Setting=Initial2025.07 | 3.53 | 40 | |
| VanillaBackbone=Llama2-7b-chat, MFT Setting=10-shot2025.07 | 3.58 | 50 | |
| VanillaBackbone=Llama2-7b, MFT Setting=50-shot2025.07 | 3.63 | 46.7 | |
| TARBackbone=Llama2-7b-chat, MFT Setting=50-shot2025.07 | 3.7 | 33.3 | |
| SimPOBackbone=Llama2-7b, MFT Setting=50-shot2025.07 | 3.8 | 46.7 | |
| T-VaccineBackbone=Llama2-7b, MFT Setting=50-shot2025.07 | 3.86 | 46.7 | |
| BoosterBackbone=Llama2-7b, MFT Setting=100-shot2025.07 | 3.86 | 53.3 | |
| VanillaBackbone=Llama2-7b, MFT Setting=10-shot2025.07 | 3.97 | 56.7 | |
| DeepAlignBackbone=Llama2-7b, MFT Setting=50-shot2025.07 | 3.97 | 42.4 | |
| DeepAlignBackbone=Llama2-7b, MFT Setting=100-shot2025.07 | 3.97 | 42.4 | |
| DeepAlignBackbone=Llama2-7b, MFT Setting=10-shot2025.07 | 4 | 50 | |
| SimPOBackbone=Llama2-7b, MFT Setting=10-shot2025.07 | 4.03 | 50 | |
| BoosterBackbone=Llama2-7b, MFT Setting=10-shot2025.07 | 4.13 | 66.7 | |
| BoosterBackbone=Llama2-7b, MFT Setting=50-shot2025.07 | 4.17 | 60 | |
| SimPOBackbone=Llama2-7b, MFT Setting=100-shot2025.07 | 4.23 | 63.3 | |
| TARBackbone=Llama2-7b-chat, MFT Setting=10-shot2025.07 | 4.3 | 56.7 | |
| TARBackbone=Llama2-7b-chat, MFT Setting=100-shot2025.07 | 4.3 | 60 | |
| VanillaBackbone=Llama2-7b-chat, MFT Setting=50-shot2025.07 | 4.52 | 80.3 | |
| VanillaBackbone=Llama2-7b-chat, MFT Setting=100-shot2025.07 | 4.54 | 80 |