Machine Translation on FLORES-200 DE-EN (test)
31.2BLEUMix-MoE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Mix-MoEArchitecture Strategy=Mix-MoE2026.05 | 31.2 | 55.8 | 86.5 | |
| MOE-LPRArchitecture Strategy=MOE-LPR2026.05 | 29.8 | 55.2 | 86.1 | |
| MoEArchitecture Strategy=MoE2026.05 | 27.5 | 48.5 | 79.2 | |
| Dense PT+SFTTraining Strategy=Single-stage Fine-tuning baseline2026.05 | 26.2 | 48.5 | 77.5 | |
| Qwen2.5-1.5BModel Scale=1.5B2026.05 | 25.6 | 42.1 | 70.5 | |
| Full FinetuneTraining Strategy=Full Finetune2026.05 | 24.5 | 46.1 | 74.2 | |
| LLAMA-ProTraining Strategy=LLAMA-Pro2026.05 | 23.1 | 38.2 | 68.4 | |
| LoRATraining Strategy=LoRA2026.05 | 22.8 | 39.5 | 69.8 | |
| LLAMA-MOE v2Architecture Strategy=LLAMA-MOE v22026.05 | 22 | 38.5 | 64.1 | |
| Qwen2.5-0.5BModel Scale=0.5B2026.05 | 20.4 | 35.8 | 64.2 | |
| Llama3.2-1BModel Scale=1B2026.05 | 18.2 | 32.4 | 61.5 |