Machine Unlearning on WMDP Cyber (Refusal Rate)
59.6Refusal RateGuardrail Filter
Evaluation Results
| Method | Links | |
|---|---|---|
| Guardrail FilterBase Model=Qwen3-14B2026.05 | 59.6 | |
| Guardrail FilterBase Model=Llama-3-8B-Instruct2026.05 | 63.7 | |
| ICCUBase Model=Qwen3-14B2026.05 | 94.3 | |
| ICCUBase Model=Llama-3-8B-Instruct2026.05 | 97.1 |