Machine Unlearning on WMDP-bio (forget-set)
0.803BF16 ScoreZero-shot
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Zero-shotBackbone=Qwen-3-8B2026.05 | 0.803 | — | — | |
| Zero-shotBackbone=Llama-3.1-8B-Instruct2026.05 | 0.763 | — | — | |
| LUNARBackbone=Qwen-3-8B2026.05 | 0.658 | 0.671 | 0.013 | |
| LUNARBackbone=Llama-3.1-8B-Instruct2026.05 | 0.621 | 0.638 | 0.017 | |
| MANSUBackbone=Qwen-3-8B2026.05 | 0.617 | 0.581 | -0.036 | |
| Surgical GABackbone=Llama-3.1-8B-Instruct2026.05 | 0.547 | 0.573 | 0.027 | |
| NPOBackbone=Llama-3.1-8B-Instruct2026.05 | 0.443 | 0.423 | -0.02 | |
| MANSUBackbone=Llama-3.1-8B-Instruct2026.05 | 0.43 | 0.39 | -0.04 | |
| NPOBackbone=Qwen-3-8B2026.05 | 0.283 | 0.32 | 0.037 | |
| GU+SimNPOBackbone=Qwen-3-8B2026.05 | 0.267 | 0.263 | -0.003 | |
| Global GABackbone=Llama-3.1-8B-Instruct2026.05 | 0.26 | 0.31 | 0.05 | |
| Surgical GABackbone=Qwen-3-8B2026.05 | 0.26 | 0.247 | -0.013 | |
| SimNPOBackbone=Llama-3.1-8B-Instruct2026.05 | 0.25 | 0.25 | 0 | |
| Global GABackbone=Qwen-3-8B2026.05 | 0.233 | 0.233 | 0 | |
| GU+SimNPOBackbone=Llama-3.1-8B-Instruct2026.05 | 0.23 | 0.23 | 0 | |
| SimNPOBackbone=Qwen-3-8B2026.05 | 0.227 | 0.227 | 0 |