Unlearning on TOFU Llama 2 1.0 (forget 10%)
85.14UFEJoint
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| JointLoss=DPO+GD (targeted)2026.05 | 85.14 | 49.94 | 65.05 | 66.3 | |
| DualOptim 8bitLoss=DPO+GD (targeted)2026.05 | 82.18 | 40.49 | 73.31 | 67.32 | |
| AlternateLoss=DPO+GD (targeted)2026.05 | 80.91 | 36.59 | 74.03 | 66.39 | |
| DualOptimLoss=DPO+GD (targeted)2026.05 | 80.77 | 38.55 | 72.43 | 66.04 | |
| DualOptim+ 8bitLoss=DPO+GD (targeted)2026.05 | 80.15 | 45.71 | 75.1 | 69.17 | |
| DualOptim+Loss=DPO+GD (targeted)2026.05 | 78.43 | 47.55 | 77.19 | 70.09 | |
| DualOptim+ 8bitLoss=NPO+GD (untargeted)2026.05 | 73.79 | — | 76.03 | 74.92 | |
| AlternateLoss=NPO+GD (untargeted)2026.05 | 72.77 | — | 75.9 | 74.34 | |
| DualOptim 8bitLoss=NPO+GD (untargeted)2026.05 | 72.42 | — | 75.75 | 74.09 | |
| DualOptim+Loss=NPO+GD (untargeted)2026.05 | 71.95 | — | 73.95 | 72.95 | |
| DualOptimLoss=NPO+GD (untargeted)2026.05 | 71.78 | — | 73.99 | 72.89 | |
| JointLoss=NPO+GD (untargeted)2026.05 | 67.9 | — | 64.46 | 66.18 |