Machine Unlearning on TOFU Llama-3.2 Instruct (Avg 1%, 5%, 10% Forget Sets)
-81.257FQ ScoreGA
Evaluation results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| GAModel Size=3B, Approach=Training-based2026.01 | -81.257 | 38.3 | 12.5 | 33.1 | 59.3 | |
| GAModel Size=1B, Approach=Training-based2026.01 | -81.114 | 19.9 | 8.6 | 24.4 | 48.4 | |
| FullModel Size=3B, Approach=Reference2026.01 | -13.96 | 66.6 | 89.9 | 88.4 | 86.8 | |
| GDModel Size=3B, Approach=Training-based2026.01 | -13.959 | 58.9 | 19.2 | 43.7 | 67.7 | |
| FullModel Size=1B, Approach=Reference2026.01 | -11.808 | 59.9 | 72.6 | 73.7 | 87.1 | |
| GDModel Size=1B, Approach=Training-based2026.01 | -8.72 | 49.1 | 11.2 | 29.5 | 78.9 | |
| TVModel Size=1B, Approach=Arithmetic-based2026.01 | -6.174 | 49.5 | 5.9 | 20.7 | 91.4 | |
| NPO+Model Size=3B, Approach=Training-based2026.01 | -5.413 | 58.7 | 14.7 | 41.5 | 89.8 | |
| TVModel Size=3B, Approach=Arithmetic-based2026.01 | -5.284 | 61.2 | 5.8 | 30.4 | 92.1 | |
| NPOModel Size=1B, Approach=Training-based2026.01 | -4.842 | 19.8 | 8.6 | 24.6 | 59.2 | |
| NPOModel Size=3B, Approach=Training-based2026.01 | -4.508 | 38 | 12.3 | 33.4 | 63.7 | |
| NPO+Model Size=1B, Approach=Training-based2026.01 | -3.528 | 49.3 | 12.2 | 31.6 | 91.1 | |
| PerTA-fisherModel Size=3B, Approach=Arithmetic-based2026.01 | -1.211 | 66.5 | 9.2 | 61.3 | 89.5 | |
| PerTA-fisherModel Size=1B, Approach=Arithmetic-based2026.01 | -0.867 | 56.2 | 8 | 41.4 | 90.8 | |
| PerTA-gradModel Size=1B, Approach=Arithmetic-based2026.01 | -0.686 | 55.6 | 7.2 | 37.6 | 91.5 | |
| PerTA-gradModel Size=3B, Approach=Arithmetic-based2026.01 | -0.669 | 66.4 | 8.2 | 56.3 | 91.3 | |
| GTModel Size=1B, Approach=Reference2026.01 | 0 | 59.6 | 6.4 | 74.8 | 89.4 | |
| GTModel Size=3B, Approach=Reference2026.01 | 0 | 65.7 | 6.6 | 88.7 | 88.7 |