Machine Translation on FLORES-200 EN-CS (test)
12.9BLEUMix-MoE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Mix-MoEArchitecture Strategy=Mix-MoE2026.05 | 12.9 | 25.8 | 85.8 | |
| MOE-LPRArchitecture Strategy=MOE-LPR2026.05 | 10.2 | 21.8 | 83.5 | |
| MoEArchitecture Strategy=MoE2026.05 | 10 | 19.5 | 82.8 | |
| Qwen2.5-1.5BModel Scale=1.5B2026.05 | 8.5 | 16.8 | 53.5 | |
| Dense PT+SFTTraining Strategy=Single-stage Fine-tuning baseline2026.05 | 8.1 | 15.2 | 55.8 | |
| Qwen2.5-0.5BModel Scale=0.5B2026.05 | 5.8 | 9.2 | 42.5 | |
| Full FinetuneTraining Strategy=Full Finetune2026.05 | 5.7 | 12.5 | 50.5 | |
| Llama3.2-1BModel Scale=1B2026.05 | 4 | 5.8 | 45.2 | |
| LLAMA-ProTraining Strategy=LLAMA-Pro2026.05 | 3.8 | 6.2 | 43.2 | |
| LoRATraining Strategy=LoRA2026.05 | 3.1 | 4.5 | 43.5 | |
| LLAMA-MOE v2Architecture Strategy=LLAMA-MOE v22026.05 | 2.6 | 3.8 | 36.1 |