Harmful Content Detection on BeaverTails Harmful (held-out target labels)
0.793AUROCQuotient Transfer
Evaluation Results
| Method | Links | |
|---|---|---|
| Quotient TransferTarget Model=Qwen-2.5-Coder-7B, Source Model=Qwen-2.5-7B-Instruct2026.05 | 0.793 | |
| Safety Monitor (In-model)Model=Qwen-2.5-7B-Instruct2026.05 | 0.791 | |
| Quotient TransferTarget Model=Qwen-14B, Source Model=Qwen-2.5-7B-Instruct2026.05 | 0.781 | |
| Quotient TransferTarget Model=Mistral, Source Model=Qwen-2.5-7B-Instruct2026.05 | 0.773 | |
| Quotient TransferTarget Model=Qwen-3B, Source Model=Qwen-2.5-7B-Instruct2026.05 | 0.772 | |
| Random-initialization control2026.05 | 0.632 |