Machine Unlearning on TOFU (forget 10% data)
97.66UFE ScoreJoint
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| JointBackbone=Llama 2, Loss Function=ME+GD (untargeted)2026.05 | 97.66 | — | 60.63 | 79.15 | |
| AlternateBackbone=Llama 2, Loss Function=ME+GD (untargeted)2026.05 | 96.86 | — | 75.34 | 86.1 | |
| DualOptim+Backbone=Llama 2, Loss Function=ME+GD (untargeted)2026.05 | 96.85 | — | 75.86 | 86.35 | |
| DualOptimBackbone=Llama 2, Loss Function=ME+GD (untargeted)2026.05 | 96.78 | — | 75.6 | 86.19 | |
| DualOptim+ 8bitBackbone=Llama 2, Loss Function=ME+GD (untargeted), Optimizer Precision=8-bit2026.05 | 96.78 | — | 75.52 | 86.15 | |
| DualOptim 8bitBackbone=Llama 2, Loss Function=ME+GD (untargeted), Optimizer Precision=8-bit2026.05 | 96.57 | — | 75.44 | 86.01 | |
| DualOptim+ 8bitBackbone=Phi 1.5, Loss Function=ME+GD (untargeted), Optimizer Precision=8-bit2026.05 | 92.78 | — | 49.96 | 71.37 | |
| DualOptim+Backbone=Phi 1.5, Loss Function=ME+GD (untargeted)2026.05 | 92.46 | — | 50.32 | 71.39 | |
| DualOptimBackbone=Phi 1.5, Loss Function=ME+GD (untargeted)2026.05 | 92.39 | — | 49.23 | 70.81 | |
| DualOptim 8bitBackbone=Phi 1.5, Loss Function=ME+GD (untargeted), Optimizer Precision=8-bit2026.05 | 92.32 | — | 48.04 | 70.19 | |
| AlternateBackbone=Phi 1.5, Loss Function=ME+GD (untargeted)2026.05 | 91.96 | — | 48.48 | 70.22 | |
| JointBackbone=Phi 1.5, Loss Function=ME+GD (untargeted)2026.05 | 91.1 | — | 36.88 | 63.99 | |
| JointBackbone=Llama 2, Loss Function=IDK+GD (targeted)2026.05 | 78.08 | 71.25 | 57.15 | 65.91 | |
| DualOptimBackbone=Llama 2, Loss Function=IDK+GD (targeted)2026.05 | 74.62 | 68.15 | 73.15 | 72.27 | |
| DualOptim 8bitBackbone=Llama 2, Loss Function=IDK+GD (targeted), Optimizer Precision=8-bit2026.05 | 74.62 | 68.15 | 73.38 | 72.38 | |
| AlternateBackbone=Llama 2, Loss Function=IDK+GD (targeted)2026.05 | 74.16 | 68.15 | 73.98 | 72.57 | |
| DualOptim+Backbone=Llama 2, Loss Function=IDK+GD (targeted)2026.05 | 73.91 | 68.15 | 75.46 | 73.25 | |
| DualOptim+ 8bitBackbone=Llama 2, Loss Function=IDK+GD (targeted), Optimizer Precision=8-bit2026.05 | 73.49 | 68.15 | 76.14 | 73.48 | |
| JointBackbone=Phi 1.5, Loss Function=IDK+GD (targeted)2026.05 | 71.65 | 64.39 | 33.92 | 50.97 | |
| DualOptimBackbone=Phi 1.5, Loss Function=IDK+GD (targeted)2026.05 | 65.87 | 66.84 | 50.25 | 58.3 | |
| AlternateBackbone=Phi 1.5, Loss Function=IDK+GD (targeted)2026.05 | 65.82 | 64.46 | 49.54 | 57.34 | |
| DualOptim 8bitBackbone=Phi 1.5, Loss Function=IDK+GD (targeted), Optimizer Precision=8-bit2026.05 | 65.81 | 65.6 | 50.38 | 58.05 | |
| DualOptim+Backbone=Phi 1.5, Loss Function=IDK+GD (targeted)2026.05 | 65.42 | 66.5 | 51.32 | 58.64 | |
| DualOptim+ 8bitBackbone=Phi 1.5, Loss Function=IDK+GD (targeted), Optimizer Precision=8-bit2026.05 | 65.26 | 65.59 | 51.3 | 58.36 |