Text Generation on OpenWebText (test)
3.77Average PerplexityBeam/Greedy Decoding
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Beam/Greedy DecodingBackbone=Llama-2-70B2024.05 | 3.77 | — | |
| Superposed DecodingBackbone=Llama-2-70B, Draft #=Best2024.05 | 4.63 | — | |
| Superposed DecodingBackbone=Llama-2-70B, Draft #=12024.05 | 5.03 | — | |
| Nucleus SamplingBackbone=Llama-2-70B2024.05 | 5.17 | — | |
| Superposed DecodingBackbone=Llama-2-70B, Draft #=22024.05 | 7.97 | — | |
| Superposed DecodingBackbone=Llama-2-70B, Draft #=32024.05 | 10.05 | — | |
| Data2026.05 | 14.8 | 5.44 | |
| ARModel Type=Autoregressive2024.12 | 17.54 | — | |
| DUO+DCDN_disc=322026.05 | 46.3 | 5.38 | |
| MDLM+SDTTN_disc=322026.05 | 62.3 | 5.49 | |
| DiLaDiffNumber of continuous steps (N_cont)=5, Number of discrete steps (N_disc)=32, Temperature (zeta)=0.72026.05 | 76.3 | 5.39 | |
| DiLaDiffNumber of continuous steps (N_cont)=5, Number of discrete steps (N_disc)=322026.05 | 108.7 | 5.53 | |
| N-Gram InterpolationBackbone=Llama-2-70B2024.05 | 152.75 | — |