Machine Unlearning on TOFU Forget 5%
64Aggregation MetricRetrain
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| RetrainBackbone=Llama 3.2 1B2025.10 | 64 | 58 | 72 | |
| RetrainBackbone=Llama 3.1 8B2025.10 | 62 | 57 | 67 | |
| RetrainBackbone=Llama 3.2 3B2025.10 | 61 | 55 | 69 | |
| BS-SBackbone=Llama 3.2 3B2025.10 | 60 | 55 | 65 | |
| BS-SBackbone=Llama 3.1 8B2025.10 | 60 | 53 | 70 | |
| BS-SBackbone=Llama 3.2 1B2025.10 | 58 | 54 | 63 | |
| BS-TBackbone=Llama 3.1 8B2025.10 | 58 | 51 | 67 | |
| NPOBackbone=Llama 3.2 3B2025.10 | 57 | 55 | 60 | |
| RMUBackbone=Llama 3.2 1B2025.10 | 55 | 49 | 63 | |
| BS-TBackbone=Llama 3.2 1B2025.10 | 55 | 53 | 57 | |
| BS-TBackbone=Llama 3.2 3B2025.10 | 55 | 53 | 62 | |
| NPOBackbone=Llama 3.2 1B2025.10 | 54 | 53 | 55 | |
| RMUBackbone=Llama 3.1 8B2025.10 | 54 | 45 | 68 | |
| WGABackbone=Llama 3.2 1B2025.10 | 53 | 45 | 64 | |
| NPOBackbone=Llama 3.1 8B2025.10 | 53 | 49 | 57 | |
| GradDiffBackbone=Llama 3.2 1B2025.10 | 52 | 48 | 57 | |
| RMUBackbone=Llama 3.2 3B2025.10 | 50 | 38 | 74 | |
| WGABackbone=Llama 3.2 3B2025.10 | 50 | 39 | 69 | |
| GradDiffBackbone=Llama 3.2 3B2025.10 | 49 | 42 | 59 | |
| GradDiffBackbone=Llama 3.1 8B2025.10 | 49 | 40 | 62 | |
| WGABackbone=Llama 3.1 8B2025.10 | 49 | 37 | 74 | |
| SimNPOBackbone=Llama 3.2 1B2025.10 | 43 | 31 | 71 | |
| SimNPOBackbone=Llama 3.2 3B2025.10 | 40 | 27 | 75 | |
| SimNPOBackbone=Llama 3.1 8B2025.10 | 36 | 24 | 70 | |
| OriginalBackbone=Llama 3.2 1B2025.10 | 16 | 9 | 71 | |
| OriginalBackbone=Llama 3.2 3B2025.10 | 6 | 3 | 75 | |
| OriginalBackbone=Llama 3.1 8B2025.10 | 2 | 1 | 73 |