Language detoxification on Bot-Adversarial Dialogue (BAD) 1.0 (test)
0.084Toxicity ProbabilityCLICK
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| CLICKbase_model=BlenderBot 365M2023.06 | 0.084 | 6.48 | 0.49 | 0.56 | |
| CLICK (Random)base_model=BlenderBot 365M, ablation_type=Random2023.06 | 0.105 | 6.48 | 0.49 | 0.56 | |
| CLICK (- Lower)base_model=BlenderBot 365M, ablation_type=Lower2023.06 | 0.105 | 6.87 | 0.5 | 0.57 | |
| Directorbase_model=BlenderBot 365M2023.06 | 0.164 | 7.93 | 0.26 | 0.29 | |
| GeDibase_model=BlenderBot 365M2023.06 | 0.187 | 7.1 | 0.14 | 0.15 | |
| CLICK (- Lowest)base_model=BlenderBot 365M, ablation_type=Lowest2023.06 | 0.2 | 7.37 | 0.52 | 0.59 | |
| DExpertsbase_model=BlenderBot 365M2023.06 | 0.303 | 8.92 | 0.4 | 0.44 | |
| Cringebase_model=BlenderBot 365M2023.06 | 0.437 | 9.06 | 0.42 | 0.49 | |
| Non-toxic FTbase_model=BlenderBot 365M2023.06 | 0.45 | 5.23 | 0.4 | 0.46 | |
| Unlikelihoodbase_model=BlenderBot 365M2023.06 | 0.453 | 6.32 | 0.42 | 0.49 |