Machine Unlearning on TOFU (forget 5% data)
97.65UFEJoint
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| JointBackbone=Llama 2, Loss Function=ME+GD (untargeted)2026.05 | 97.65 | — | 57.15 | 77.4 | |
| AlternateBackbone=Llama 2, Loss Function=ME+GD (untargeted)2026.05 | 97.07 | — | 75.64 | 86.36 | |
| DualOptim+Backbone=Llama 2, Loss Function=ME+GD (untargeted)2026.05 | 96.91 | — | 76.09 | 86.5 | |
| DualOptim 8bitBackbone=Llama 2, Loss Function=ME+GD (untargeted), Optimizer Precision=8-bit2026.05 | 96.74 | — | 75.52 | 86.13 | |
| DualOptim+ 8bitBackbone=Llama 2, Loss Function=ME+GD (untargeted), Optimizer Precision=8-bit2026.05 | 96.69 | — | 76.16 | 86.43 | |
| DualOptimBackbone=Llama 2, Loss Function=ME+GD (untargeted)2026.05 | 96.66 | — | 75.9 | 86.28 | |
| DualOptim+ 8bitBackbone=Phi 1.5, Loss Function=ME+GD (untargeted), Optimizer Precision=8-bit2026.05 | 93.13 | — | 50.55 | 71.84 | |
| DualOptim 8bitBackbone=Phi 1.5, Loss Function=ME+GD (untargeted), Optimizer Precision=8-bit2026.05 | 93.12 | — | 50.99 | 72.06 | |
| DualOptim+Backbone=Phi 1.5, Loss Function=ME+GD (untargeted)2026.05 | 93.07 | — | 50.87 | 71.97 | |
| AlternateBackbone=Phi 1.5, Loss Function=ME+GD (untargeted)2026.05 | 92.3 | — | 49.73 | 71.02 | |
| DualOptimBackbone=Phi 1.5, Loss Function=ME+GD (untargeted)2026.05 | 91.97 | — | 51.73 | 71.86 | |
| JointBackbone=Phi 1.5, Loss Function=ME+GD (untargeted)2026.05 | 91.32 | — | 33.87 | 62.6 | |
| JointBackbone=Llama 2, Loss Function=IDK+GD (targeted)2026.05 | 80.29 | 68.6 | 55.63 | 65.04 | |
| DualOptimBackbone=Llama 2, Loss Function=IDK+GD (targeted)2026.05 | 76.58 | 70.27 | 73.64 | 73.53 | |
| DualOptim 8bitBackbone=Llama 2, Loss Function=IDK+GD (targeted), Optimizer Precision=8-bit2026.05 | 76.42 | 70.27 | 73.71 | 73.53 | |
| AlternateBackbone=Llama 2, Loss Function=IDK+GD (targeted)2026.05 | 75.79 | 70.27 | 73.93 | 73.48 | |
| DualOptim+Backbone=Llama 2, Loss Function=IDK+GD (targeted)2026.05 | 75.11 | 70.27 | 75.91 | 74.3 | |
| DualOptim+ 8bitBackbone=Llama 2, Loss Function=IDK+GD (targeted), Optimizer Precision=8-bit2026.05 | 75.04 | 70.27 | 76.27 | 74.42 | |
| JointBackbone=Phi 1.5, Loss Function=IDK+GD (targeted)2026.05 | 72.55 | 58.32 | 36.26 | 50.85 | |
| DualOptimBackbone=Phi 1.5, Loss Function=IDK+GD (targeted)2026.05 | 68.49 | 64.34 | 49.5 | 57.96 | |
| DualOptim 8bitBackbone=Phi 1.5, Loss Function=IDK+GD (targeted), Optimizer Precision=8-bit2026.05 | 68.34 | 64.33 | 48.41 | 57.37 | |
| AlternateBackbone=Phi 1.5, Loss Function=IDK+GD (targeted)2026.05 | 67.73 | 64.3 | 47.81 | 56.91 | |
| DualOptim+Backbone=Phi 1.5, Loss Function=IDK+GD (targeted)2026.05 | 67.63 | 67.6 | 51.52 | 59.57 | |
| DualOptim+ 8bitBackbone=Phi 1.5, Loss Function=IDK+GD (targeted), Optimizer Precision=8-bit2026.05 | 67.56 | 65.94 | 50.36 | 58.55 |