Machine Unlearning on LLaMA-3-8B Unlearning Evaluation Suite (test)
71.09AccuracyOrigin
Evaluation Results
| Method | Links | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| OriginBackbone=LLaMA-3-8B2026.05 | 71.09 | 5.92 | 94.12 | 12.14 | 2.36 | 23.67 | 100 | 86.2 | 100 | 63.82 | 1,720 | |
| IDK+APBackbone=LLaMA-3-8B2026.05 | 70.7 | 15.01 | 35.46 | 9.77 | 1.86 | 11.32 | 100 | 95.2 | 100 | 63.59 | 1,704 | |
| SimNPOBackbone=LLaMA-3-8B2026.05 | 28.44 | 0 | 0 | 57.17 | 57.23 | 49.46 | — | 0.6 | 0.8 | 55.79 | 39 | |
| GraddiffBackbone=LLaMA-3-8B2026.05 | 27.81 | 0.21 | 12.5 | 0.9 | 0.82 | 2.18 | — | 83.6 | 89.8 | 46.15 | 505 | |
| NPOBackbone=LLaMA-3-8B2026.05 | 27.65 | 0.21 | 0 | 71.01 | 71.12 | 44.78 | — | 6.4 | 76.4 | 50.58 | 519 | |
| NPO+SAMBackbone=LLaMA-3-8B2026.05 | 26.39 | 0 | 0 | 57.14 | 57.14 | 49.49 | — | 0 | 0 | 51.09 | 2 |