Text Generation on Wikitext (Human Evaluation)
88.7Coherence: CD Better RateContrastive Decoding
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Contrastive DecodingCD Model=GPT-2 XL, Baseline Decoding Strategy=typical, Baseline Model=GPT-2 XL2022.10 | 88.7 | 4.6 | 6.7 | 70.3 | 8.2 | 21.5 | |
| Contrastive DecodingCD Model=OPT-13B, Baseline Decoding Strategy=typical, Baseline Model=OPT-13B2022.10 | 77.3 | 10.6 | 12.1 | 68.7 | 15.2 | 16.2 | |
| Contrastive DecodingCD Model=GPT-2 XL, Baseline Decoding Strategy=nucleus, Baseline Model=GPT-2 XL2022.10 | 71.4 | 8.3 | 20.2 | 54.8 | 8.3 | 36.9 | |
| Contrastive DecodingCD Model=OPT-13B, Baseline Decoding Strategy=nucleus, Baseline Model=OPT-13B2022.10 | 55.6 | 20.2 | 24.2 | 41.9 | 19.7 | 38.4 |