Unconditional Generation on Unconditional Generations 1024
6.8PPL (GPT-NEO)GPT2-XL
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| GPT2-XLSampling Algorithm=top-p, p = 0.8, L = 1024, T = 1024, Total Params=1.6B2026.05 | 6.8 | — | — | |
| GPT2-LSampling Algorithm=top-p, p = 0.8, L = 1024, T = 1024, Total Params=774M2026.05 | 7.4 | — | 6.5 | |
| Glauber-UL2-LConfiguration=post-SEDD, T=3 × L, i.e., N = 3, Sampling Algorithm=L = 1024, Total Params=898M2026.05 | 7.8 | — | 6.9 | |
| Glauber-UL2-LConfiguration=post-SEDD, T=L, i.e., N = 1, Sampling Algorithm=L = 1024, Total Params=898M2026.05 | 9.9 | — | 9.5 | |
| UL2-LConfiguration=post-SEDD, T=3 × L CAUSAL-GEN, Sampling Algorithm=L = 1024, Total Params=898M2026.05 | 11.5 | — | 11.4 | |
| GPT2-MSampling Algorithm=top-p, p = 0.8, L = 1024, T = 1024, Total Params=345M2026.05 | 14.5 | 12.4 | 13 | |
| Glauber-UL2-MConfiguration=post-SEDD, T=3 × L, i.e., N = 3, Sampling Algorithm=L = 1024, Total Params=419M2026.05 | 14.9 | 13.2 | 13.7 | |
| UL2-LConfiguration=pre-SEDD CAUSAL-GEN, Sampling Algorithm=L = 1024, Total Params=783M2026.05 | 14.9 | — | 14.2 | |
| Glauber-UL2-MConfiguration=post-SEDD, T=L, i.e., N = 1, Sampling Algorithm=L = 1024, Total Params=419M2026.05 | 16.6 | 17.1 | 17.5 | |
| PlaidSampling Algorithm=τ = 0.9, L = 1024, T = 4096, Total Params=1.3B2026.05 | 17.9 | 19.7 | 19.7 | |
| GGMSampling Algorithm=top-p, p = 0.8, L = 1024, T = 4096, Total Params=387M2026.05 | 18 | 19.5 | 19.9 | |
| UL2-MConfiguration=post-SEDD, T=3 × L CAUSAL-GEN, Sampling Algorithm=L = 1024, Total Params=419M2026.05 | 19.9 | 19.1 | 19.6 | |
| UL2-MConfiguration=pre-SEDD CAUSAL-GEN, Sampling Algorithm=L = 1024, Total Params=368M2026.05 | 20.4 | 21.7 | 21.3 | |
| SEDD-MSampling Algorithm=L = 1024, T = 2048, Total Params=424M2026.05 | 25.2 | 27.3 | 28 | |
| MDLMSampling Algorithm=L = 1024, T = 1000, Total Params=170M2026.05 | 40.9 | 44.2 | 45.4 |