Machine Unlearning on TOFU forget05 1.0
75.85Model Utility (MU)Base
Evaluation Results
| Method | Links | |||||
|---|---|---|---|---|---|---|
| Base2026.03 | 75.85 | 3.19 | 39.52 | — | — | |
| IDKVariant=AP2026.03 | 75.23 | 60.88 | 68.05 | — | — | |
| ME_KLTarget Model=Llama-3.1-8B, Method category=Convergence-based2026.03 | 75.14 | 70.15 | 72.65 | — | — | |
| MEVariant=GD2026.03 | 74.96 | 70.15 | 72.56 | — | — | |
| MEVariant=KL2026.03 | 74.43 | 70.44 | 72.43 | — | — | |
| ASUVariant=KL2026.03 | 74.18 | 77.84 | 76.01 | — | — | |
| ASUVariant=GD2026.03 | 73.62 | 77.58 | 75.6 | — | — | |
| IDK_APTarget Model=Llama-3.1-8B, Method category=Convergence-based2026.03 | 72.74 | 60.93 | 66.83 | — | — | |
| ASU_KLTarget Model=Llama-3.1-8B2026.03 | 71.67 | 74.07 | 72.87 | — | — | |
| RMU (KLR)Model=Llama-3.2-3B-Instruct2026.05 | 67 | — | 16 | 9 | 87 | |
| SimNPO (KLR)Model=Llama-3.2-3B-Instruct2026.05 | 66 | — | 48 | 37 | 88 | |
| GUARD-ITModel=Llama-3.2-3B-Instruct2026.05 | 66 | — | 61 | 57 | 83 | |
| SatImpModel=Llama-3.2-3B-Instruct2026.05 | 64 | — | 48 | 38 | 76 | |
| SimNPOModel=Llama-3.2-3B-Instruct2026.05 | 64 | — | 39 | 28 | 88 | |
| GradDiffModel=Llama-3.2-3B-Instruct2026.05 | 62 | — | 76 | 100 | 29 | |
| WGAModel=Llama-3.2-3B-Instruct2026.05 | 61 | — | 47 | 38 | 88 | |
| GUARD-ITModel=Llama-3.2-1B-Instruct2026.05 | 60 | — | 62 | 63 | 81 | |
| RMUModel=Llama-3.2-3B-Instruct2026.05 | 60 | — | 24 | 15 | 88 | |
| NPO_KLTarget Model=Llama-3.1-8B, Method category=Divergence-based2026.03 | 59.99 | 60.78 | 60.38 | — | — | |
| SimNPO (KLR)Model=Llama-3.2-1B-Instruct2026.05 | 59 | — | 45 | 37 | 88 | |
| RMU (KLR)Model=Llama-3.2-1B-Instruct2026.05 | 58 | — | 17 | 10 | 90 | |
| NPOVariant=KL2026.03 | 57.7 | 73.35 | 65.52 | — | — | |
| SatImpModel=Llama-3.2-1B-Instruct2026.05 | 57 | — | 46 | 39 | 93 | |
| SimNPOModel=Llama-3.2-1B-Instruct2026.05 | 57 | — | 42 | 33 | 91 | |
| NPOVariant=GD2026.03 | 56.62 | 73.31 | 64.97 | — | — | |
| WGAModel=Llama-3.2-1B-Instruct2026.05 | 56 | — | 49 | 43 | 89 | |
| RMUModel=Llama-3.2-1B-Instruct2026.05 | 55 | — | 29 | 20 | 89 | |
| UNDIALModel=Llama-3.2-3B-Instruct2026.05 | 51 | — | 57 | 64 | 83 | |
| UNDIALModel=Llama-3.2-1B-Instruct2026.05 | 45 | — | 52 | 62 | 87 | |
| GA_KLTarget Model=Llama-3.1-8B, Method category=Divergence-based2026.03 | 36.34 | 0 | 18.17 | — | — | |
| NPOModel=Llama-3.2-3B-Instruct2026.05 | 34 | — | 39 | 46 | 90 | |
| GradDiffModel=Llama-3.2-1B-Instruct2026.05 | 32 | — | 46 | 82 | 23 | |
| NPO (KLR)Model=Llama-3.2-3B-Instruct2026.05 | 31 | — | 37 | 46 | 87 | |
| GAVariant=GD2026.03 | 29.25 | 3.89 | 16.57 | — | — | |
| GAVariant=KL2026.03 | 20.13 | 5.39 | 12.76 | — | — | |
| NPO (KLR)Model=Llama-3.2-1B-Instruct2026.05 | 3 | — | 5 | 42 | 87 | |
| DPOModel=Llama-3.2-3B-Instruct2026.05 | 3 | — | 5 | 34 | 64 | |
| DPOModel=Llama-3.2-1B-Instruct2026.05 | 2 | — | 4 | 34 | 93 | |
| NPOModel=Llama-3.2-1B-Instruct2026.05 | 2 | — | 3 | 37 | 85 | |
| DPO_KLTarget Model=Llama-3.1-8B, Method category=Convergence-based2026.03 | 1.74 | 68.51 | 35.12 | — | — | |
| DPOVariant=GD2026.03 | 0 | 77.15 | 38.58 | — | — | |
| DPOVariant=KL2026.03 | 0 | 77.22 | 38.61 | — | — | |
| IDKVariant=GD2026.03 | 0 | 70.18 | 35.09 | — | — | |
| IDKVariant=KL2026.03 | 0 | 70.18 | 35.09 | — | — | |
| IDK_KLTarget Model=Llama-3.1-8B, Method category=Convergence-based2026.03 | 0 | 64.68 | 32.34 | — | — | |
| CEUModel=Llama-3.2-1B-Instruct2026.05 | 0 | — | 0 | 24 | 4 | |
| GAModel=Llama-3.2-1B-Instruct2026.05 | 0 | — | 0 | 97 | 10 | |
| GradDiff (KLR)Model=Llama-3.2-1B-Instruct2026.05 | 0 | — | 0 | 69 | 43 | |
| PDUModel=Llama-3.2-1B-Instruct2026.05 | 0 | — | 0 | 14 | 19 | |
| CEUModel=Llama-3.2-3B-Instruct2026.05 | 0 | — | 0 | 27 | 3 | |
| GAModel=Llama-3.2-3B-Instruct2026.05 | 0 | — | 0 | 100 | 5 | |
| GradDiff (KLR)Model=Llama-3.2-3B-Instruct2026.05 | 0 | — | 0 | 98 | 8 | |
| PDUModel=Llama-3.2-3B-Instruct2026.05 | 0 | — | 0 | 4 | 5 |