Machine Translation on FLORES-200 TR-EN (test)
19.1BLEUMix-MoE
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Mix-MoEArchitecture Strategy=Mix-MoE2026.05 | 19.1 | 38.8 | 73.2 | |
| MOE-LPRArchitecture Strategy=MOE-LPR2026.05 | 18 | 37.1 | 72.1 | |
| MoEArchitecture Strategy=MoE2026.05 | 17.2 | 33.8 | 67.2 | |
| Dense PT+SFTTraining Strategy=Single-stage Fine-tuning baseline2026.05 | 16.8 | 37.1 | 72.5 | |
| Full FinetuneTraining Strategy=Full Finetune2026.05 | 16.2 | 36.5 | 72.1 | |
| LLAMA-ProTraining Strategy=LLAMA-Pro2026.05 | 15 | 30.5 | 66.5 | |
| Qwen2.5-1.5BModel Scale=1.5B2026.05 | 14.5 | 26.5 | 64.5 | |
| LoRATraining Strategy=LoRA2026.05 | 14.5 | 28.2 | 66.1 | |
| LLAMA-MOE v2Architecture Strategy=LLAMA-MOE v22026.05 | 13.8 | 22.8 | 65.1 | |
| Qwen2.5-0.5BModel Scale=0.5B2026.05 | 11.2 | 21 | 58.9 | |
| Llama3.2-1BModel Scale=1B2026.05 | 9.8 | 18.5 | 55.4 |