Unlearning on TOFU Phi 1.5 1.0 (forget 1%)
74.62UFEAlternate
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| AlternateLoss=NPO+GD (untargeted)2026.05 | 74.62 | — | 48.32 | 61.47 | |
| DualOptimLoss=NPO+GD (untargeted)2026.05 | 74.88 | — | 48.47 | 61.68 | |
| DualOptim 8bitLoss=NPO+GD (untargeted)2026.05 | 75.51 | — | 48.77 | 62.14 | |
| DualOptim+Loss=NPO+GD (untargeted)2026.05 | 75.8 | — | 48.72 | 62.26 | |
| DualOptim+ 8bitLoss=NPO+GD (untargeted)2026.05 | 76.1 | — | 48.79 | 62.45 | |
| AlternateLoss=DPO+GD (targeted)2026.05 | 78.56 | 35.11 | 48.23 | 52.53 | |
| JointLoss=NPO+GD (untargeted)2026.05 | 78.87 | — | 29.89 | 54.38 | |
| JointLoss=DPO+GD (targeted)2026.05 | 79.43 | 23.74 | 32.08 | 41.83 | |
| DualOptimLoss=DPO+GD (targeted)2026.05 | 81.58 | 55.36 | 46.32 | 57.39 | |
| DualOptim+Loss=DPO+GD (targeted)2026.05 | 82.92 | 61.9 | 48.05 | 60.23 | |
| DualOptim+ 8bitLoss=DPO+GD (targeted)2026.05 | 83.46 | 57.44 | 47.61 | 59.03 | |
| DualOptim 8bitLoss=DPO+GD (targeted)2026.05 | 83.77 | 57.56 | 47.73 | 59.2 |