Machine Translation on FLORES-200 EN-TR (test)
12.1BLEUMix-MoE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Mix-MoEArchitecture Strategy=Mix-MoE2026.05 | 12.1 | 23.5 | 67.2 | |
| Qwen2.5-1.5BModel Scale=1.5B2026.05 | 11.5 | 16.2 | 64.1 | |
| MOE-LPRArchitecture Strategy=MOE-LPR2026.05 | 10.2 | 21 | 65.5 | |
| MoEArchitecture Strategy=MoE2026.05 | 10 | 18.2 | 65.8 | |
| Qwen2.5-0.5BModel Scale=0.5B2026.05 | 8.9 | 12.5 | 52.1 | |
| Dense PT+SFTTraining Strategy=Single-stage Fine-tuning baseline2026.05 | 8.1 | 15.5 | 57.5 | |
| Llama3.2-1BModel Scale=1B2026.05 | 7.8 | 10.5 | 48.2 | |
| Full FinetuneTraining Strategy=Full Finetune2026.05 | 6.3 | 13.5 | 54.5 | |
| LLAMA-ProTraining Strategy=LLAMA-Pro2026.05 | 5.2 | 10.5 | 48.2 | |
| LLAMA-MOE v2Architecture Strategy=LLAMA-MOE v22026.05 | 5.1 | 9.2 | 52.1 | |
| LoRATraining Strategy=LoRA2026.05 | 3.9 | 5.8 | 45.8 |