Unconstrained Text Generation L=1024 (test)
15.5MAUVE (1/8 NFE)SM (iso-update)
Evaluation results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| SM (iso-update)Batch size=512, Gradient updates=1M, Forward tokens=1049B, Loss tokens=262B, Sequence length (L)=1024, Unmasking type=Standard2025.10 | 15.5 | 38.3 | 53.5 | 60.2 | 39.6 | 30.7 | 26.1 | 23.5 | |
| SM (iso-compute)Batch size=512, Gradient updates=0.5M, Forward tokens=524B, Loss tokens=131B, Sequence length (L)=1024, Unmasking type=Standard2025.10 | 14.3 | 41.7 | 49.8 | 59.6 | 41.1 | 32 | 27.4 | 24.6 | |
| MDLMBatch size=512, Gradient updates=1M, Forward tokens=524B, Loss tokens=262B, Sequence length (L)=1024, Unmasking type=Standard2025.10 | 1.7 | 2.5 | 3.6 | 3.4 | 60 | 55 | 52.4 | 50.5 | |
| CADDBatch size=256, Gradient updates=2M, Forward tokens=524B, Loss tokens=262B, Sequence length (L)=1024, Unmasking type=Standard2025.10 | 1.7 | 8 | 14 | 24 | 55.3 | 49.6 | 46.1 | 44.6 |