Harmful Prompt Detection on AdvBench & XSTest (held-out evaluation sets)
0.9642AUROC (Harmful/Normal)LatentBiopsy
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| LatentBiopsyModel=Qwen3.5-0.8B (D=1024), Type=Base, Layer=202026.03 | 0.9642 | 1 | 0.9373 | 1 | 0.9758 | 0.384 | 92.8 | |
| LatentBiopsyModel=Qwen2.5-0.5B (D=896), Type=Base, Layer=202026.03 | 0.9585 | 1 | 0.9373 | 1 | 0.972 | 0.149 | 90.2 | |
| LatentBiopsyModel=Qwen3.5-0.8B (D=1024), Type=Abliterated, Layer=202026.03 | 0.9517 | 1 | 0.9165 | 1 | 0.9674 | 0.427 | 89.9 | |
| LatentBiopsyModel=Qwen3.5-0.8B (D=1024), Type=Instruct, Layer=202026.03 | 0.9497 | 1 | 0.9117 | 1 | 0.9661 | 0.434 | 89.9 | |
| LatentBiopsyModel=Qwen2.5-0.5B (D=896), Type=Instruct, Layer=202026.03 | 0.942 | 1 | 0.9129 | 1 | 0.9609 | 0.219 | 87.5 | |
| LatentBiopsyModel=Qwen2.5-0.5B (D=896), Type=Abliterated, Layer=102026.03 | 0.9374 | 1 | 0.8978 | 1 | 0.9577 | 0.179 | 88.2 |