OOD Detection on ToxicChat (test)
60.2Length-Matched AUROCBest Single
Evaluation Results
| Method | Links | |
|---|---|---|
| Best Singlecategory=Trajectory, backbone=Qwen3-0.6B, evaluation_mode=length-matched, description=Best single trajectory feature of 92026.04 | 60.2 | |
| Maha-Trajcategory=Trajectory, backbone=Qwen3-0.6B, evaluation_mode=length-matched2026.04 | 58 | |
| RAUQcategory=Attn-Based, backbone=Qwen3-0.6B, evaluation_mode=length-matched2026.04 | 53 | |
| CEDcategory=Attn-Based, backbone=Qwen3-0.6B, evaluation_mode=length-matched2026.04 | 50.9 | |
| Attn Ent.category=Attn-Based, backbone=Qwen3-0.6B, evaluation_mode=length-matched, description=mean attention entropy across heads/layers2026.04 | 50.9 |