Machine Unlearning on WMDP v1 (test)
42.43Accuracy (Cyber)Origin
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| OriginModel Architecture=Qwen1.5-MoE-A2.7B-Chat2026.06 | 42.43 | 65.36 | 53.9 | 59.54 | |
| OriginModel Architecture=DeepSeek-V2-Lite-Chat2026.06 | 39.46 | 58.6 | 49.03 | 56.72 | |
| SEUFModel Architecture=DeepSeek-V2-Lite-Chat2026.06 | 30.5 | 26 | 28.25 | 45.53 | |
| ESFTModel Architecture=Qwen1.5-MoE-A2.7B-Chat2026.06 | 29.99 | 26.55 | 28.27 | 53.19 | |
| GradDiffModel Architecture=DeepSeek-V2-Lite-Chat2026.06 | 29.94 | 32.67 | 31.31 | 37.22 | |
| GAModel Architecture=Qwen1.5-MoE-A2.7B-Chat2026.06 | 29.64 | 26.47 | 28.06 | 32.17 | |
| GAModel Architecture=DeepSeek-V2-Lite-Chat2026.06 | 29.29 | 27.02 | 28.16 | 33.07 | |
| TRACEModel Architecture=Qwen1.5-MoE-A2.7B-Chat2026.06 | 27.48 | 26.63 | 27.06 | 58.19 | |
| NPOModel Architecture=Qwen1.5-MoE-A2.7B-Chat2026.06 | 26.87 | 24.74 | 25.81 | 24.81 | |
| GradDiffModel Architecture=Qwen1.5-MoE-A2.7B-Chat2026.06 | 26.57 | 24.67 | 25.62 | 22.95 | |
| TRACEModel Architecture=DeepSeek-V2-Lite-Chat2026.06 | 26.47 | 26.24 | 26.36 | 55.41 | |
| ESFTModel Architecture=DeepSeek-V2-Lite-Chat2026.06 | 25.92 | 23.96 | 24.94 | 50.99 | |
| SEUFModel Architecture=Qwen1.5-MoE-A2.7B-Chat2026.06 | 25.45 | 26.7 | 26.08 | 49.15 | |
| NPOModel Architecture=DeepSeek-V2-Lite-Chat2026.06 | 22.46 | 25.31 | 23.89 | 24.66 |