Safety Classification on BEAVERTAILS 1,000 examples native human labels (test)
85.5F1 ScorePost-hoc Qwen3Guard-Gen-8B
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Post-hoc Qwen3Guard-Gen-8BMode=Post-hoc, Backbone=Gen-8B2026.06 | 85.5 | 83.2 | 88 | 15.5 | 14 | 466 | 8.02 | 448 | 551 | 602 | |
| Streaming Qwen3Guard-Stream-8BMode=Streaming, Backbone=Stream-8B2026.06 | 84.6 | 82.9 | 86.4 | 14.2 | 0 | 3,089 | 1.32 | 39 | 43 | 54 | |
| MLP-TOKEN ProbeOperating point=per-task threshold calibrated to maximise F12026.06 | 79.5 | 79.8 | 79.2 | 8 | 0 | 7.2 | 0.6 | 0.1 | 0.1 | 0.1 | |
| MLP-TOKEN ProbeOperating point=default threshold (zt ≥ 0)2026.06 | 50.4 | 93 | 34.6 | 8 | 0 | 7.2 | 0.6 | 0.1 | 0.1 | 0.1 |