Red Teaming on ConvAI2 (filtered hard positive)
2,120RSRBRT (e+r)
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| BRT (e+r)Target Model=BB-3B, Query limit (NQ)=200002023.05 | 2,120 | 33 | |
| SLTarget Model=BB-3B, Query limit (NQ)=20000, Backbone=OPT-1.3B2023.05 | 780 | 52 | |
| SFSTarget Model=BB-3B, Query limit (NQ)=20000, Backbone=OPT-1.3B2023.05 | 370 | 44.8 | |
| SFSTarget Model=BB-3B, Query limit (NQ)=20000, Backbone=Bloom2023.05 | 360 | 44.7 | |
| Offensive Top-NQTarget Model=BB-3B, Query limit (NQ)=200002023.05 | 340 | 35.8 | |
| BRT (s+r)Target Model=BB-3B, Query limit (NQ)=200002023.05 | 340 | 34.7 | |
| RandTarget Model=BB-3B, Query limit (NQ)=200002023.05 | 80 | 36.3 |