Detoxification on REALTOXICITYPROMPTS (test)
0.081Toxicity Score (Avg)FINE-GRAINED RLHF
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| FINE-GRAINED RLHFReward Type=Sentence-level (fine-grained)2023.06 | 0.081 | 9.77 | 94.9 | 93.2 | |
| Holistic RLHFReward Type=Holistic2023.06 | 0.13 | 11.75 | 94.3 | 92.6 | |
| DEXPERTSApproach=Controlled Generation2023.06 | 0.136 | 22.83 | 93.2 | 92.2 | |
| GeDiApproach=Controlled Generation2023.06 | 0.154 | 24.78 | 93.8 | 93.8 | |
| GPT-2Model Size=Large2023.06 | 0.192 | 9.58 | 94.7 | 93.1 |