Toxicity Detection on Safety Evaluation Set Toxicity (held-out target labels)
97.6AUROCSafety Monitor (In-model)
Evaluation Results
| Method | Links | |
|---|---|---|
| Safety Monitor (In-model)Model=Qwen-2.5-7B-Instruct2026.05 | 97.6 | |
| Quotient TransferTarget Model=Qwen-2.5-Coder-7B, Source Model=Qwen-2.5-7B-Instruct, Protocol=Zero-label alignment2026.05 | 97.1 | |
| Quotient TransferTarget Model=Qwen-14B, Source Model=Qwen-2.5-7B-Instruct, Protocol=Zero-label alignment2026.05 | 95.5 | |
| Quotient TransferTarget Model=Qwen-3B, Source Model=Qwen-2.5-7B-Instruct, Protocol=Zero-label alignment2026.05 | 95.2 | |
| Quotient TransferTarget Model=Mistral, Source Model=Qwen-2.5-7B-Instruct, Protocol=Zero-label alignment2026.05 | 92.6 | |
| Random-initialization control2026.05 | 58 |