Multilingual Machine Translation (X→En) on WMT (test)
43BLEU (De)STABLEMOE
Evaluation Results
| Method | Links | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| STABLEMOENumber of Parameters=480M, Training FLOPs=317B2022.04 | 43 | 37.4 | 34.7 | 31.5 | 29.3 | 26.8 | 24.7 | 23.6 | 21.9 | 30.32 | |
| Hash LayerNumber of Parameters=480M, Training FLOPs=317B2022.04 | 42.7 | 37 | 34.6 | 31.3 | 28.7 | 26.5 | 23.9 | 23.1 | 21.7 | 29.94 | |
| BASE LayerNumber of Parameters=480M, Training FLOPs=317B2022.04 | 42.6 | 37.8 | 34.2 | 31 | 29 | 26.9 | 25.1 | 23.2 | 21.6 | 30.16 | |
| Switch TransformerNumber of Parameters=480M, Training FLOPs=317B2022.04 | 42.3 | 37.1 | 33.8 | 31 | 28.6 | 26 | 24.3 | 23 | 21.2 | 29.7 | |
| Larger TransformerNumber of Parameters=90M, Training FLOPs=317B2022.04 | 40.6 | 36.9 | 33.7 | 29.8 | 27.8 | 25.4 | 24.6 | 22.2 | 20.9 | 29.1 | |
| Standard TransformerNumber of Parameters=77M, Training FLOPs=290B2022.04 | 39.8 | 36 | 32.5 | 29.1 | 27.2 | 24.5 | 23.6 | 21.8 | 20.3 | 28.31 |