Machine Translation on FLORES-200 EN-RU (test)
13.1BLEU ScoreMix-MoE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Mix-MoEArchitecture Strategy=Mix-MoE2026.05 | 13.1 | 27.8 | 77.8 | |
| Qwen2.5-1.5BModel Scale=1.5B2026.05 | 12.1 | 25.8 | 74.2 | |
| MOE-LPRArchitecture Strategy=MOE-LPR2026.05 | 11.8 | 24.2 | 76.5 | |
| Dense PT+SFTTraining Strategy=Single-stage Fine-tuning baseline2026.05 | 11.4 | 22.5 | 67.5 | |
| Qwen2.5-0.5BModel Scale=0.5B2026.05 | 11.3 | 23.5 | 69.8 | |
| MoEArchitecture Strategy=MoE2026.05 | 10.9 | 22.5 | 68.1 | |
| Full FinetuneTraining Strategy=Full Finetune2026.05 | 9.5 | 20.8 | 64.8 | |
| LLAMA-ProTraining Strategy=LLAMA-Pro2026.05 | 4.9 | 8.2 | 47.5 | |
| LoRATraining Strategy=LoRA2026.05 | 4.4 | 8.5 | 45.5 | |
| LLAMA-MOE v2Architecture Strategy=LLAMA-MOE v22026.05 | 3.9 | 8.1 | 47.1 | |
| Llama3.2-1BModel Scale=1B2026.05 | 3.8 | 5.5 | 41.5 |