Detoxification on Jigsaw (test)
20.8Perplexity (PPL)Prefix
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| PrefixControl Optimization Strategy=Optimization in the Latent Space2022.12 | 20.8 | 88.3 | 16.3 | 43.8 | 67.5 | — | — | — | — | |
| GPT2(raw)status=raw2022.12 | 25.06 | — | — | — | — | 10,397.67 | 17.1 | 52.71 | 37.13 | |
| T5Training Mode=Finetune LM2022.12 | 27.21 | — | — | — | — | 42.04 | 22.81 | 39.83 | 63.49 | |
| PFCategory=Lightweight methods2022.12 | 28.67 | — | — | — | — | 52.73 | 38.37 | 49.68 | 41.53 | |
| GPT2 + PT + PL + noiseBackbone=GPT2, Training Mode=Self-Training, Components=PT, PL, noise2022.12 | 29.2 | — | — | — | — | 26.37 | 40.99 | 49.75 | 43.1 | |
| Ctr-PFCategory=Lightweight methods2022.12 | 29.28 | — | — | — | — | 57.39 | 31.53 | 49.47 | 46.7 | |
| GPT2 + PT(pos) + PLBackbone=GPT2, Training Mode=Self-Training, Components=PT(pos), PL2022.12 | 29.49 | — | — | — | — | 25.87 | 40 | 49.52 | 43.22 | |
| GPT2 + PT(select) + PLBackbone=GPT2, Training Mode=Self-Training, Components=PT(select), PL2022.12 | 29.83 | — | — | — | — | 26.44 | 43.45 | 49.65 | 43.03 | |
| GPT2Training Mode=Finetune LM2022.12 | 32.79 | — | — | — | — | 66.61 | 43.94 | 51.62 | 42.05 | |
| GPT2 + PT + noiseBackbone=GPT2, Training Mode=Self-Training, Components=PT, noise2022.12 | 34.69 | — | — | — | — | 66.12 | 40.59 | 51.31 | 43.42 | |
| GPT2 + PTBackbone=GPT2, Training Mode=Self-Training, Components=Pseudo-training (PT)2022.12 | 36.29 | — | — | — | — | 71.47 | 41.03 | 51.91 | 42.15 | |
| Con. PrefixControl Optimization Strategy=Optimization in the Latent Space2022.12 | 37.7 | 93.8 | 17.3 | 47 | 71.1 | — | — | — | — | |
| UniLM + PT(select) + PLBackbone=UniLM, Training Mode=Self-Training, Components=PT(select), PL2022.12 | 40.7 | — | — | — | — | 54.5 | 29.21 | 45.42 | 46.94 | |
| UniLM + PT + PL + noiseBackbone=UniLM, Training Mode=Self-Training, Components=PT, PL, noise2022.12 | 40.98 | — | — | — | — | 55.99 | 26.95 | 44.47 | 47.07 | |
| UniLM + PT(pos) + PLBackbone=UniLM, Training Mode=Self-Training, Components=PT(pos), PL2022.12 | 45.09 | — | — | — | — | 55.87 | 25.13 | 45.91 | 46.7 | |
| DiscreteControl Optimization Strategy=Optimization in the Latent Space2022.12 | 46.2 | 90.1 | 36.9 | 76.3 | 87 | — | — | — | — | |
| UniLM + PTBackbone=UniLM, Training Mode=Self-Training, Components=Pseudo-training (PT)2022.12 | 46.78 | — | — | — | — | 74.71 | 34.68 | 36.82 | 55.89 | |
| UniLM + PT + noiseBackbone=UniLM, Training Mode=Self-Training, Components=PT, noise2022.12 | 51.99 | — | — | — | — | 80.46 | 39.46 | 40.16 | 52.95 | |
| UniLMTraining Mode=Finetune LM2022.12 | 52.23 | — | — | — | — | 67.92 | 34.38 | 38.26 | 55.31 | |
| PriorControlControl Optimization Strategy=Optimization in the Latent Space2022.12 | 54.3 | 90.7 | 29.1 | 70.1 | 86.9 | — | — | — | — | |
| PriorControl + extendControl Optimization Strategy=Optimization in the Latent Space, Method Variant=Extended2022.12 | 54.6 | 95.7 | 29.8 | 70.5 | 86.8 | — | — | — | — | |
| DuNST-PTDescription=DuNST without pseudo text, only using pseudo-labeled data2022.12 | 56.75 | — | — | — | — | 50.28 | 15.32 | 47.03 | 47.1 | |
| LatentOpsControl Optimization Strategy=Optimization in the Latent Space2022.12 | 58.8 | 94.6 | 13.5 | 48.3 | 62.8 | — | — | — | — | |
| PPLMControl Optimization Strategy=Biasing during Decoding2022.12 | 63.2 | 93.2 | 31.1 | 70.9 | 85.9 | — | — | — | — | |
| Mix&MatchControl Optimization Strategy=Optimization in the Language Space2022.12 | 65.2 | 96.9 | 31.5 | 74.8 | 88.8 | — | — | — | — | |
| DuNST(pos)Components=Positional information included2022.12 | 74.74 | — | — | — | — | 63.75 | 13.69 | 50.37 | 42.62 | |
| GeDiControl Optimization Strategy=Biasing during Decoding2022.12 | 81.6 | 94.9 | 38.1 | 74 | 78.4 | — | — | — | — | |
| GeDi rawControl Optimization Strategy=Biasing during Decoding, Method Variant=Raw2022.12 | 134.1 | 95.4 | 47.5 | 88.9 | 93 | — | — | — | — | |
| MUCOCOControl Optimization Strategy=Optimization in the Language Space2022.12 | 381.7 | 94.8 | 22.5 | 49.9 | 64.3 | — | — | — | — |