Unlearning on TOFU Llama 2 1.0 (forget 5%)
88.11UFE ScoreJoint
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| JointLoss=DPO+GD (targeted)2026.05 | 88.11 | 45.17 | 66.98 | 66.81 | |
| DualOptim 8bitLoss=DPO+GD (targeted)2026.05 | 84.49 | 48.19 | 73.07 | 69.71 | |
| DualOptimLoss=DPO+GD (targeted)2026.05 | 83.67 | 45.31 | 71.42 | 67.96 | |
| AlternateLoss=DPO+GD (targeted)2026.05 | 83.13 | 46.41 | 73.46 | 69.11 | |
| DualOptim+ 8bitLoss=DPO+GD (targeted)2026.05 | 81.94 | 47.83 | 75.06 | 69.98 | |
| DualOptim+Loss=DPO+GD (targeted)2026.05 | 77.99 | 41.86 | 76.51 | 68.22 | |
| DualOptim+Loss=NPO+GD (untargeted)2026.05 | 75.94 | — | 75.37 | 75.65 | |
| DualOptim+ 8bitLoss=NPO+GD (untargeted)2026.05 | 75.42 | — | 76.01 | 75.72 | |
| DualOptim 8bitLoss=NPO+GD (untargeted)2026.05 | 75.35 | — | 75.11 | 75.24 | |
| AlternateLoss=NPO+GD (untargeted)2026.05 | 75.32 | — | 75.69 | 75.51 | |
| DualOptimLoss=NPO+GD (untargeted)2026.05 | 74.31 | — | 74.52 | 74.42 | |
| JointLoss=NPO+GD (untargeted)2026.05 | 69.67 | — | 63.12 | 66.4 |