Unlearning on TOFU Phi 1.5 1.0 (forget 10%)
66.88UFEDualOptim 8bit
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| DualOptim 8bitLoss=NPO+GD (untargeted)2026.05 | 66.88 | — | 50.07 | 58.48 | |
| DualOptim+ 8bitLoss=NPO+GD (untargeted)2026.05 | 67.11 | — | 51.25 | 59.18 | |
| DualOptimLoss=NPO+GD (untargeted)2026.05 | 67.33 | — | 48.73 | 58.03 | |
| AlternateLoss=NPO+GD (untargeted)2026.05 | 67.81 | — | 49.15 | 58.48 | |
| DualOptim+Loss=NPO+GD (untargeted)2026.05 | 69.12 | — | 49.71 | 59.42 | |
| DualOptim 8bitLoss=DPO+GD (targeted)2026.05 | 72.76 | 40.95 | 50.02 | 53.44 | |
| DualOptimLoss=DPO+GD (targeted)2026.05 | 72.97 | 52.61 | 50.28 | 56.53 | |
| DualOptim+ 8bitLoss=DPO+GD (targeted)2026.05 | 73.14 | 46.21 | 50.24 | 54.96 | |
| JointLoss=NPO+GD (untargeted)2026.05 | 73.73 | — | 27.78 | 50.76 | |
| AlternateLoss=DPO+GD (targeted)2026.05 | 74.49 | 57.51 | 49.87 | 57.94 | |
| DualOptim+Loss=DPO+GD (targeted)2026.05 | 74.67 | 56.46 | 50.78 | 58.18 | |
| JointLoss=DPO+GD (targeted)2026.05 | 77.67 | 45.46 | 31.48 | 46.52 |