Language Generation on WebText (completions)
10.16Perplexity (PPL)Coherence Tuning
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Coherence Tuningtuning_mode=self-tune, base_method=Nucleus (p=0.95)2021.10 | 10.16 | 0.33 | 0.95 | 64 | 16.19 | 9.85 | 6.59 | 4.16 | |
| SamplingTemperature=0.9, source=Holtzman et al. (2019)2021.10 | 10.6 | 0.35 | 0.96 | 66 | 16.36 | 10.01 | 6.54 | 4.15 | |
| Coherence Boostingk=32, alpha_k=-0.05, base_method=Nucleus (p=0.95)2021.10 | 12.81 | 0.31 | 0.94 | 34 | 15.54 | 9.42 | 6.16 | 3.98 | |
| Coherence Boostingk=64, alpha_k=-0.1, base_method=Nucleus (p=0.95)2021.10 | 12.93 | 0.32 | 0.95 | 46 | 15.75 | 9.67 | 6.1 | 3.95 | |
| Human2021.10 | 13.19 | 0.31 | 0.93 | 28 | 15.95 | 9.51 | 6.54 | 4.03 | |
| Nucleus Samplingp=0.95, source=Holtzman et al. (2019)2021.10 | 13.48 | 0.32 | 0.95 | 46 | 15.06 | 9.11 | 5.65 | 3.62 | |
| Samplingsource=Holtzman et al. (2019)2021.10 | 23.53 | 0.28 | 0.93 | 22 | 12.92 | 7.71 | 4.87 | 3.28 |