Safety Evaluation on DarkBench (test)
95.5User RetentionQwen-8b
Evaluation results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| Qwen-8bModel=Qwen-8b2026.01 | 95.5 | 37.3 | 24.5 | 7.3 | 15.5 | 17.3 | 0.329 | |
| GPT-OSS-20bModel=GPT-OSS-20b2026.01 | 82.7 | 13.6 | 50 | 6.4 | 1.8 | 33.9 | 0.314 | |
| Gemma-12bModel=Gemma-12b2026.01 | 70.9 | 76.4 | 33.6 | 24.5 | 11.8 | 8.2 | 0.376 | |
| NSED-R3Method=NSED, Deliberation Round=32026.01 | 52.2 | 72.4 | 25 | 14.3 | 15.6 | 4.8 | 0.307 | |
| NSED-R1Method=NSED, Deliberation Round=12026.01 | 42.9 | 74.1 | 35.3 | 11.1 | 15 | 12.5 | 0.318 | |
| NSED-R2Method=NSED, Deliberation Round=22026.01 | 25 | 81 | 28 | 4 | 9.1 | 24.1 | 0.285 |