Machine Unlearning on WMDP Bio (Metric: Refusal Rate)
39.5Refusal RateGuardrail Filter
Evaluation Results
| Method | Links | |
|---|---|---|
| Guardrail FilterBase Model=Llama-3-8B-Instruct2026.05 | 39.5 | |
| Guardrail FilterBase Model=Qwen3-14B2026.05 | 47 | |
| ICCUBase Model=Qwen3-14B2026.05 | 96 | |
| ICCUBase Model=Llama-3-8B-Instruct2026.05 | 96.8 |