THUMT is an open-source toolkit for neural machine translation developed by the Tsinghua Natural Language Processing Group.
We benchmark THUMT on the following datasets:
The testsets for WMT14 En-De and WMT18 Zh-En are newstest2014
and newstest2017
respectively.
Dataset | Model | Size | Steps | GPUs | Batch/GPU | Mode | BLEU |
---|---|---|---|---|---|---|---|
WMT14 En-De | Transformer | Base | 100k | 4 | 2*4096 | FP16 | 26.85 |
WMT14 En-De | Transformer | Base | 100k | 4 | 2*4096 | FP32 | 26.91 |
WMT14 En-De | Transformer | Base | 100k | 8 | 4096 | FP32 | 26.95 |
WMT14 En-De | Transformer | Base | 86k | 8 | 2*4096 | FP32 | 27.21 |
WMT14 En-De | Transformer | Big | 300k | 8 | 4096 | FP16 | 28.71 |
WMT14 En-De | Transformer | Big | 20k | 16 | 8*4096 | DistFP16 | 28.68 |
WMT18 Zh-En | Transformer | Big | 300k | 8 | 2*4096 | FP16 | 24.07 |