Controlled Text Generation on RealToxicityPrompts 10K nontoxic prompts
30.2Avg Max ToxicityDEXPERTS
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| DEXPERTSExpert Size=small, alpha=22021.05 | 30.2 | 11.8 | 38.2 | 56 | 82 | 83 | |
| DEXPERTSExpert Size=medium, alpha=22021.05 | 30.7 | 12.5 | 32.51 | 57 | 84 | 84 | |
| DEXPERTSExpert Size=large, alpha=22021.05 | 31.4 | 12.8 | 32.41 | 58 | 84 | 84 | |
| DEXPERTSConfiguration=anti-only, alpha=22021.05 | 35.2 | 19.1 | 52.02 | 58 | 80 | 73 | |
| GeDiSampling=Nucleus sampling (p=0.9)2021.05 | 36.3 | 21.7 | 60.03 | 62 | 84 | 83 | |
| DAPTSampling=Nucleus sampling (p=0.9)2021.05 | 42.8 | 36 | 31.21 | 57 | 84 | 84 | |
| Non-toxic expertBase Model=GPT-2 Large, Sampling=Nucleus sampling (p=0.9)2021.05 | 48.5 | 46.4 | 40.61 | 58 | 86 | 86 | |
| PPLM (10%)Evaluation subset=1K prompts, Sampling=Nucleus sampling (p=0.9)2021.05 | 52 | 51.8 | 32.58 | 58 | 86 | 86 | |
| GPT-2Sampling=Nucleus sampling (p=0.9)2021.05 | 52.7 | 52 | 25.45 | 58 | 85 | 85 |