Machine Translation on FLORES-200 EN-ET (test)
13.1BLEUMix-MoE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Mix-MoEArchitecture Strategy=Mix-MoE2026.05 | 13.1 | 25.8 | 64.2 | |
| MoEArchitecture Strategy=MoE2026.05 | 10.9 | 17.5 | 57.8 | |
| MOE-LPRArchitecture Strategy=MOE-LPR2026.05 | 10.7 | 20.8 | 61.8 | |
| Qwen2.5-1.5BModel Scale=1.5B2026.05 | 9.2 | 11.5 | 42.5 | |
| Dense PT+SFTTraining Strategy=Single-stage Fine-tuning baseline2026.05 | 7.7 | 14.8 | 51.5 | |
| Llama3.2-1BModel Scale=1B2026.05 | 7.5 | 9.5 | 39.5 | |
| Qwen2.5-0.5BModel Scale=0.5B2026.05 | 6.8 | 8.5 | 38.1 | |
| Full FinetuneTraining Strategy=Full Finetune2026.05 | 5.9 | 12.5 | 47.8 | |
| LLAMA-ProTraining Strategy=LLAMA-Pro2026.05 | 4.8 | 8.2 | 43.5 | |
| LLAMA-MOE v2Architecture Strategy=LLAMA-MOE v22026.05 | 4.3 | 5.2 | 44.8 | |
| LoRATraining Strategy=LoRA2026.05 | 4.1 | 6.5 | 40.2 |