Harmful Knowledge Removal on WMDP Bio
78.5Acc_rBase
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| BaseBackbone=Llama-3-8B2026.04 | 78.5 | 75.3 | 5.9 | — | |
| OracleBackbone=Llama-3-8B2026.04 | 77.37 | — | 6.1 | — | |
| RMUBackbone=Llama-3-8B2026.04 | 74.63 | 0 | 7.1 | 12.5 | |
| FLATBackbone=Llama-3-8B2026.04 | 73.92 | 0.01 | 8.36 | 12.13 | |
| GABackbone=Llama-3-8B2026.04 | 73.27 | 0 | 11.27 | 12.25 | |
| STAMP-LRBackbone=Llama-3-8B2026.04 | 73.27 | 0 | 7 | 4.25 | |
| NPOBackbone=Llama-3-8B2026.04 | 71.37 | 0 | 9.27 | 11.17 | |
| WGABackbone=Llama-3-8B2026.04 | 70.17 | 2.1 | 11.99 | 11.2 | |
| STAMPBackbone=Llama-3-8B2026.04 | 70.13 | 0 | 6.55 | 7.13 | |
| ASUBackbone=Llama-3-8B2026.04 | 68.39 | 0.9 | 7.91 | 12.13 |