Dangerous Knowledge Unlearning on WMDP
43S-unlearning Scorecosine-based meta-learned RMU
Evaluation results
| Method | Links | |
|---|---|---|
| cosine-based meta-learned RMUBase Model=Mistral-7B2026.05 | 43 | |
| RMUBase Model=Mistral-7B2026.05 | 40 | |
| cosine-based meta-learned RMUBase Model=Olmo3-7B2026.05 | 35 | |
| cosine-based meta-learned RMUBase Model=Qwen2.5-7B2026.05 | 33 | |
| Adaptive RMUBase Model=Llama3.1-8B2026.05 | 32 | |
| BaseBase Model=Qwen2.5-7B2026.05 | 30 | |
| RMUBase Model=Qwen2.5-7B2026.05 | 30 | |
| Adaptive RMUBase Model=Qwen2.5-7B2026.05 | 30 | |
| BaseBase Model=Llama3.1-8B2026.05 | 29 | |
| RMUBase Model=Llama3.1-8B2026.05 | 29 | |
| BaseBase Model=Mistral-7B2026.05 | 28 | |
| RMUBase Model=Olmo3-7B2026.05 | 28 | |
| Adaptive RMUBase Model=Olmo3-7B2026.05 | 28 | |
| BaseBase Model=Olmo3-7B2026.05 | 27 | |
| cosine-based meta-learned RMUBase Model=Llama3.1-8B2026.05 | 0 | |
| Adaptive RMUBase Model=Mistral-7B2026.05 | 0 |