Harmful Refusal on Sorry-Bench
85.5Refusal Rate (String Match)LlamaGuard prompt-filter
Evaluation Results
| Method | Links | |
|---|---|---|
| LlamaGuard prompt-filterDraft Model=Llama-3.1-8B2026.06 | 85.5 | |
| LlamaGuard response-filterDraft Model=Llama-3.1-8B2026.06 | 83.4 | |
| Hard prefixDraft Model=Llama-3.1-8B, K (Beam Count)=12026.06 | 63 | |
| ALIGNBEAMDraft Model=Llama-3.1-8B, K (Beam Count)=32026.06 | 61.4 | |
| Safety promptDraft Model=Llama-3.1-8B2026.06 | 15.2 | |
| Base draftDraft Model=Llama-3.1-8B2026.06 | 10.9 |