Safety Evaluation on ADVERSARIALQA
43.75Chinese AccuracySFT+DARE
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| SFT+DAREFine-tuning protocol=PEFT2024.02 | 43.75 | 6.28 | 34.56 | 39.75 | 14.35 | 27.74 | 26.23 | |
| SFTFine-tuning protocol=PEFT2024.02 | 41.35 | 10.28 | 31.28 | 41.1 | 24.71 | 29.74 | 28.23 | |
| SFTFine-tuning protocol=Full-FT2024.02 | 31.29 | 1.57 | 20 | 23.84 | 5.73 | 12.03 | 10.52 | |
| SFT+DAREFine-tuning protocol=Full-FT2024.02 | 30.62 | 1.58 | 19.39 | 27.91 | 6.91 | 12.02 | 10.51 | |
| SFT+RESTAFine-tuning protocol=PEFT2024.02 | 18.18 | 1.07 | 10.16 | 18.53 | 3.44 | 10.28 | 8.77 | |
| SFT+RESTAdFine-tuning protocol=PEFT2024.02 | 11.97 | 0 | 10.73 | 16.38 | 3.84 | 8.58 | 7.07 | |
| SFT+RESTAFine-tuning protocol=Full-FT2024.02 | 7.79 | 0.52 | 3.61 | 5.43 | 0.51 | 2.59 | 1.08 | |
| SFT+RESTAdFine-tuning protocol=Full-FT2024.02 | 7.74 | 0 | 4.14 | 6.08 | 0.51 | 2.48 | 0.97 |