Open-ended Text Generation on Wikitext (test)
95Diversity (DIV)Typical Sampling
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Typical SamplingModel=GPT2-XL, Decoding Strategy=typical=0.952022.10 | 95 | 84 | 53 | |
| Nucleus SamplingModel=OPT-13B, Decoding Strategy=p=0.952022.10 | 92 | 89 | 55 | |
| Nucleus SamplingModel=GPT2-XL, Decoding Strategy=p=0.952022.10 | 92 | 87 | 57 | |
| Contrastive DecodingModel=OPT-13B, Decoding Strategy=CD2022.10 | 91 | 91 | 69 | |
| Typical SamplingModel=OPT-13B, Decoding Strategy=typical=0.952022.10 | 89 | 86 | 58 | |
| Contrastive DecodingModel=GPT2-XL, Decoding Strategy=CD2022.10 | 89 | 92 | 69 | |
| Contrastive SearchModel=OPT-13B, Decoding Strategy=CS2022.10 | 87 | 77 | 52 | |
| Top-k SamplingModel=GPT2-XL, Decoding Strategy=k=502022.10 | 87 | 79 | 61 | |
| Contrastive SearchModel=GPT2-XL, Decoding Strategy=CS2022.10 | 86 | 75 | 59 | |
| Top-k SamplingModel=OPT-13B, Decoding Strategy=k=502022.10 | 72 | 77 | 64 | |
| Greedy DecodingModel=OPT-13B, Decoding Strategy=max prob2022.10 | 3 | 8 | 63 | |
| Greedy DecodingModel=GPT2-XL, Decoding Strategy=max prob2022.10 | 2 | 5 | 62 |