Machine Unlearning on TOFU 1% (forget set)
-2.17FQFull
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| FullModel Size=1B, Method Category=Reference2026.01 | -2.17 | 59.9 | 74.3 | 73.7 | 89.4 | |
| NPOModel Size=1B, Method Category=Training-based2026.01 | -2.062 | 59.5 | 17.8 | 65 | 90.4 | |
| GAModel Size=1B, Method Category=Training-based2026.01 | -1.953 | 59.7 | 18.9 | 65.6 | 90.9 | |
| GDModel Size=1B, Method Category=Training-based2026.01 | -1.845 | 58.1 | 16.9 | 56.2 | 90.7 | |
| NPO+Model Size=1B, Method Category=Training-based2026.01 | -1.845 | 59.6 | 17.4 | 65.6 | 90.7 | |
| FullModel Size=3B, Method Category=Reference2026.01 | -1.845 | 66.6 | 92 | 88.4 | 89.4 | |
| GAModel Size=3B, Method Category=Training-based2026.01 | -1.845 | 66.8 | 25.2 | 82.4 | 86.4 | |
| GDModel Size=3B, Method Category=Training-based2026.01 | -1.845 | 66.3 | 32 | 82.6 | 89.7 | |
| NPOModel Size=3B, Method Category=Training-based2026.01 | -1.845 | 66.8 | 25.3 | 82.5 | 83.8 | |
| NPO+Model Size=3B, Method Category=Training-based2026.01 | -1.845 | 66.9 | 25.4 | 81.9 | 85.6 | |
| PerTA-fisherModel Size=1B, Method Category=Task Arithmetic-based2026.01 | -0.576 | 58.6 | 8.5 | 60 | 89.5 | |
| TVModel Size=1B, Method Category=Task Arithmetic-based2026.01 | -0.393 | 55.6 | 8.1 | 35.8 | 90.8 | |
| PerTA-gradModel Size=1B, Method Category=Task Arithmetic-based2026.01 | -0.289 | 58.1 | 7.5 | 55.1 | 91.2 | |
| TVModel Size=3B, Method Category=Task Arithmetic-based2026.01 | -0.238 | 65.6 | 7.5 | 55 | 93.3 | |
| PerTA-fisherModel Size=3B, Method Category=Task Arithmetic-based2026.01 | -0.238 | 67.2 | 10.6 | 80.3 | 86.9 | |
| PerTA-gradModel Size=3B, Method Category=Task Arithmetic-based2026.01 | -0.037 | 66.9 | 8.5 | 75.7 | 90.3 | |
| GTModel Size=1B, Method Category=Reference2026.01 | 0 | 59.9 | 6.9 | 75.1 | 87.4 | |
| GTModel Size=3B, Method Category=Reference2026.01 | 0 | 66.2 | 6.7 | 88.8 | 90.4 |