Machine Translation on FLORES-200 EN-DE (test)
23.5BLEUMix-MoE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Mix-MoEArchitecture Strategy=Mix-MoE2026.05 | 23.5 | 44.2 | 73.8 | |
| Dense PT+SFTTraining Strategy=Single-stage Fine-tuning baseline2026.05 | 18 | 28.5 | 60.5 | |
| MOE-LPRArchitecture Strategy=MOE-LPR2026.05 | 17.1 | 35.1 | 67.8 | |
| Qwen2.5-1.5BModel Scale=1.5B2026.05 | 16.8 | 35.2 | 71.5 | |
| MoEArchitecture Strategy=MoE2026.05 | 15.8 | 31.2 | 70.2 | |
| Qwen2.5-0.5BModel Scale=0.5B2026.05 | 12.5 | 25.5 | 61.2 | |
| Full FinetuneTraining Strategy=Full Finetune2026.05 | 12.1 | 25.2 | 56.8 | |
| Llama3.2-1BModel Scale=1B2026.05 | 11 | 16.5 | 53.5 | |
| LLAMA-ProTraining Strategy=LLAMA-Pro2026.05 | 6.8 | 13.5 | 46.8 | |
| LLAMA-MOE v2Architecture Strategy=LLAMA-MOE v22026.05 | 5.4 | 9.8 | 36.2 | |
| LoRATraining Strategy=LoRA2026.05 | 3.3 | 6.2 | 45.1 |