Machine Unlearning on TOFU 10% Forget Set (v1.0)
-238.973FQGA
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| GAModel Size=1B, Category=Training-based2026.01 | -238.973 | 0 | 3.3 | 3.5 | 0.125 | |
| GAModel Size=3B, Category=Training-based2026.01 | -236.07 | 0 | 3.3 | 3.5 | 0.05 | |
| GDModel Size=3B, Category=Training-based2026.01 | -26.8 | 55.3 | 11.7 | 24.2 | 0.556 | |
| FullModel Size=3B, Status=Before unlearning2026.01 | -26.444 | 66.6 | 89 | 88.4 | 0.861 | |
| FullModel Size=1B, Status=Before unlearning2026.01 | -21.408 | 59.9 | 70.6 | 73.7 | 0.861 | |
| GDModel Size=1B, Category=Training-based2026.01 | -15.484 | 43.4 | 7.6 | 15.1 | 0.707 | |
| TVModel Size=1B, Category=Task Arithmetic-based2026.01 | -12.506 | 45.1 | 4.6 | 11.4 | 0.895 | |
| NPOModel Size=1B, Category=Training-based2026.01 | -10.244 | 0 | 3.3 | 3.5 | 0.329 | |
| TVModel Size=3B, Category=Task Arithmetic-based2026.01 | -10.22 | 55.1 | 4.8 | 15 | 0.904 | |
| NPO+Model Size=3B, Category=Training-based2026.01 | -7.042 | 54.6 | 8.7 | 22.4 | 0.926 | |
| NPOModel Size=3B, Category=Training-based2026.01 | -4.59 | 0 | 3.4 | 3.8 | 0.206 | |
| NPO+Model Size=1B, Category=Training-based2026.01 | -4.481 | 42.3 | 9.3 | 15.1 | 0.946 | |
| PerTA-fisherModel Size=3B, Category=Task Arithmetic-based2026.01 | -2.99 | 64.7 | 8.8 | 47.4 | 0.911 | |
| PerTA-gradModel Size=3B, Category=Task Arithmetic-based2026.01 | -1.708 | 64.9 | 8.2 | 43.2 | 0.921 | |
| PerTA-fisherModel Size=1B, Category=Task Arithmetic-based2026.01 | -1.686 | 54.8 | 7.7 | 29.5 | 0.919 | |
| PerTA-gradModel Size=1B, Category=Task Arithmetic-based2026.01 | -1.107 | 54.2 | 7.1 | 26.6 | 0.922 | |
| GTModel Size=1B, Status=Ground truth model2026.01 | 0 | 59.1 | 5.9 | 74.6 | 0.904 | |
| GTModel Size=3B, Status=Ground truth model2026.01 | 0 | 65 | 6.5 | 89.9 | 0.89 |