Toxic Output Mitigation on RealToxicityPrompts 1.0 (Random)
0.122ToxicityPREADD-D
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| PREADD-DBase language model=OPT-6.7B, Sampling strategy=Pure sampling, Generation length=32 tokens, Prefix type=Dynamic2023.07 | 0.122 | 56.6 | 0.284 | |
| FUDGEBase language model=OPT-6.7B, Sampling strategy=Top-k sampling (k=100), Generation length=32 tokens2023.07 | 0.128 | 22.1 | 0.329 | |
| PREADD-SBase language model=OPT-6.7B, Sampling strategy=Pure sampling, Generation length=32 tokens, Prefix type=Static2023.07 | 0.134 | 51.7 | 0.29 | |
| GBase language model=OPT-6.7B, Sampling strategy=Pure sampling, Generation length=32 tokens2023.07 | 0.152 | 49.9 | 0.301 | |
| NEGPROMPTBase language model=OPT-6.7B, Sampling strategy=Pure sampling, Generation length=32 tokens2023.07 | 0.2 | 54.3 | 0.294 |