Machine Translation on FLORES-200 RU-EN (test)
22.8BLEUMix-MoE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Mix-MoEArchitecture Strategy=Mix-MoE2026.05 | 22.8 | 44.9 | 82.5 | |
| Qwen2.5-1.5BModel Scale=1.5B2026.05 | 22.4 | 41.8 | 71.2 | |
| MOE-LPRArchitecture Strategy=MOE-LPR2026.05 | 21.8 | 43.8 | 81.8 | |
| MoEArchitecture Strategy=MoE2026.05 | 21 | 40.8 | 75.2 | |
| Dense PT+SFTTraining Strategy=Single-stage Fine-tuning baseline2026.05 | 20.5 | 41.8 | 76.8 | |
| Full FinetuneTraining Strategy=Full Finetune2026.05 | 19.2 | 40.5 | 75.1 | |
| Qwen2.5-0.5BModel Scale=0.5B2026.05 | 17.8 | 36.5 | 66.4 | |
| LLAMA-ProTraining Strategy=LLAMA-Pro2026.05 | 17.5 | 36.5 | 69.1 | |
| LoRATraining Strategy=LoRA2026.05 | 17.1 | 35.8 | 68.2 | |
| LLAMA-MOE v2Architecture Strategy=LLAMA-MOE v22026.05 | 16.5 | 33.5 | 59.2 | |
| Llama3.2-1BModel Scale=1B2026.05 | 14 | 32.1 | 62.8 |