Safety Evaluation on Safety Benchmarks VLG-1, VLG-2, Ch3EF
93.91VLG-1 ScoreSPPFT
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| SPPFTData Condition=0.1% Harmful Data Injected2025.11 | 93.91 | 80.09 | 17.86 | 63.95 | 61.09 | |
| EWCData Condition=0.1% Harmful Data Injected2025.11 | 83.15 | 47.96 | 47.23 | 59.45 | 56.58 | |
| SEFEData Condition=0.1% Harmful Data Injected2025.11 | 82.62 | 27.6 | 30.18 | 46.8 | 43.94 | |
| SeqFTData Condition=0.1% Harmful Data Injected2025.11 | 81.54 | 45.7 | 47.43 | 58.22 | 55.36 | |
| SEFEData Condition=Original Data2025.11 | 79.75 | 22.17 | 25.46 | 42.46 | 39.6 | |
| SPPFTData Condition=Original Data2025.11 | 77.96 | 47.96 | 13.14 | 46.35 | 43.49 | |
| ReplayData Condition=0.1% Harmful Data Injected2025.11 | 77.24 | 42.31 | 46.82 | 55.46 | 52.59 | |
| SeqFTData Condition=Original Data2025.11 | 76.88 | 25.34 | 25.46 | 42.56 | 39.7 | |
| EWCData Condition=Original Data2025.11 | 76.7 | 24.89 | 25.46 | 42.35 | 39.49 | |
| Model TailorData Condition=0.1% Harmful Data Injected2025.11 | 67.2 | 15.16 | 39.19 | 40.52 | 37.65 | |
| ReplayData Condition=Original Data2025.11 | 60.39 | 12.9 | 19.92 | 31.07 | 28.21 | |
| Model TailorData Condition=Original Data2025.11 | 60.22 | 7.92 | 16.73 | 28.29 | 25.43 | |
| Safe DeltaData Condition=0.1% Harmful Data Injected2025.11 | 50.9 | 5.66 | 16.22 | 24.26 | 21.4 | |
| HPAData Condition=0.1% Harmful Data Injected2025.11 | 1.43 | 5.66 | 14.58 | 7.22 | 4.36 | |
| Safe DeltaData Condition=Original Data2025.11 | 1.25 | 4.56 | 9.24 | 5.02 | 2.15 | |
| HPAData Condition=Original Data2025.11 | 0.72 | 4.3 | 9.24 | 4.75 | 1.89 | |
| Zero-shot2025.11 | 0.18 | 2.04 | 6.37 | 2.86 | — |