Machine Unlearning on TOFU (1% Forget Data)
97.88UFEDualOptim+
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| DualOptim+Backbone=Llama 2, Loss Function=ME+GD (untargeted)2026.05 | 97.88 | — | 75.82 | 86.85 | |
| DualOptim 8bitBackbone=Llama 2, Loss Function=ME+GD (untargeted), Optimizer Precision=8-bit2026.05 | 97.76 | — | 75.55 | 86.66 | |
| DualOptim+ 8bitBackbone=Llama 2, Loss Function=ME+GD (untargeted), Optimizer Precision=8-bit2026.05 | 97.5 | — | 75.01 | 86.26 | |
| DualOptimBackbone=Llama 2, Loss Function=ME+GD (untargeted)2026.05 | 97.46 | — | 75.06 | 86.26 | |
| AlternateBackbone=Llama 2, Loss Function=ME+GD (untargeted)2026.05 | 97.25 | — | 74.89 | 86.07 | |
| JointBackbone=Llama 2, Loss Function=ME+GD (untargeted)2026.05 | 95.89 | — | 59.7 | 77.8 | |
| JointBackbone=Phi 1.5, Loss Function=ME+GD (untargeted)2026.05 | 95.41 | — | 11.45 | 53.43 | |
| DualOptim+Backbone=Phi 1.5, Loss Function=ME+GD (untargeted)2026.05 | 93.92 | — | 46.19 | 70.06 | |
| DualOptim 8bitBackbone=Phi 1.5, Loss Function=ME+GD (untargeted), Optimizer Precision=8-bit2026.05 | 92.83 | — | 45.75 | 69.29 | |
| DualOptimBackbone=Phi 1.5, Loss Function=ME+GD (untargeted)2026.05 | 92.79 | — | 45.26 | 69.03 | |
| DualOptim+ 8bitBackbone=Phi 1.5, Loss Function=ME+GD (untargeted), Optimizer Precision=8-bit2026.05 | 92.48 | — | 46.16 | 69.32 | |
| AlternateBackbone=Phi 1.5, Loss Function=ME+GD (untargeted)2026.05 | 91.46 | — | 45.78 | 68.62 | |
| JointBackbone=Llama 2, Loss Function=IDK+GD (targeted)2026.05 | 85.96 | 59.5 | 38.62 | 55.68 | |
| DualOptimBackbone=Llama 2, Loss Function=IDK+GD (targeted)2026.05 | 83.12 | 66.41 | 73.83 | 74.3 | |
| DualOptim 8bitBackbone=Llama 2, Loss Function=IDK+GD (targeted), Optimizer Precision=8-bit2026.05 | 83.11 | 66.41 | 73.77 | 74.27 | |
| AlternateBackbone=Llama 2, Loss Function=IDK+GD (targeted)2026.05 | 81.97 | 66.41 | 73.93 | 74.06 | |
| DualOptim+ 8bitBackbone=Llama 2, Loss Function=IDK+GD (targeted), Optimizer Precision=8-bit2026.05 | 79.05 | 66.41 | 76.53 | 74.63 | |
| DualOptim+Backbone=Llama 2, Loss Function=IDK+GD (targeted)2026.05 | 78.76 | 66.41 | 77.4 | 74.99 | |
| JointBackbone=Phi 1.5, Loss Function=IDK+GD (targeted)2026.05 | 78.11 | 45.45 | 18.61 | 40.19 | |
| DualOptim 8bitBackbone=Phi 1.5, Loss Function=IDK+GD (targeted), Optimizer Precision=8-bit2026.05 | 75.58 | 61.23 | 46.73 | 57.57 | |
| DualOptim+Backbone=Phi 1.5, Loss Function=IDK+GD (targeted)2026.05 | 75.51 | 67.85 | 47.69 | 59.69 | |
| DualOptimBackbone=Phi 1.5, Loss Function=IDK+GD (targeted)2026.05 | 74.75 | 63.51 | 46.46 | 57.8 | |
| DualOptim+ 8bitBackbone=Phi 1.5, Loss Function=IDK+GD (targeted), Optimizer Precision=8-bit2026.05 | 73.69 | 68.36 | 47.53 | 59.28 | |
| AlternateBackbone=Phi 1.5, Loss Function=IDK+GD (targeted)2026.05 | 73.35 | 62.49 | 48.14 | 58.03 |