Unlearning on TOFU Llama 2 1.0 (forget 1%)
92.02UFEJoint
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| JointLoss=DPO+GD (targeted)2026.05 | 92.02 | 45.74 | 44.19 | 56.54 | |
| DualOptim 8bitLoss=DPO+GD (targeted)2026.05 | 89.47 | 46.54 | 72.32 | 70.16 | |
| DualOptimLoss=DPO+GD (targeted)2026.05 | 89.26 | 48.24 | 72.33 | 70.54 | |
| AlternateLoss=DPO+GD (targeted)2026.05 | 88.27 | 53.77 | 73.67 | 72.35 | |
| DualOptim+ 8bitLoss=DPO+GD (targeted)2026.05 | 82.04 | 43.57 | 77.01 | 69.91 | |
| DualOptim+Loss=DPO+GD (targeted)2026.05 | 79.49 | 45.96 | 77.51 | 70.12 | |
| DualOptim+Loss=NPO+GD (untargeted)2026.05 | 78.51 | — | 75.07 | 76.8 | |
| JointLoss=NPO+GD (untargeted)2026.05 | 77.31 | — | 55.56 | 66.44 | |
| DualOptim+ 8bitLoss=NPO+GD (untargeted)2026.05 | 76.58 | — | 74.91 | 75.75 | |
| DualOptim 8bitLoss=NPO+GD (untargeted)2026.05 | 73.85 | — | 74.22 | 74.04 | |
| DualOptimLoss=NPO+GD (untargeted)2026.05 | 73.83 | — | 74.51 | 74.17 | |
| AlternateLoss=NPO+GD (untargeted)2026.05 | 73.51 | — | 74.4 | 73.96 |