Machine Translation on FLORES-200 FI-EN (test)
15.8BLEUMix-MoE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Mix-MoEArchitecture Strategy=Mix-MoE2026.05 | 15.8 | 35.1 | 76.1 | |
| MOE-LPRArchitecture Strategy=MOE-LPR2026.05 | 14.2 | 34.1 | 75.2 | |
| MoEArchitecture Strategy=MoE2026.05 | 13 | 31.5 | 74.2 | |
| Dense PT+SFTTraining Strategy=Single-stage Fine-tuning baseline2026.05 | 12.5 | 31.8 | 69.8 | |
| Qwen2.5-1.5BModel Scale=1.5B2026.05 | 12.1 | 23.5 | 62.4 | |
| Full FinetuneTraining Strategy=Full Finetune2026.05 | 11.8 | 30.5 | 68.1 | |
| LLAMA-ProTraining Strategy=LLAMA-Pro2026.05 | 10.9 | 27.5 | 63.5 | |
| LoRATraining Strategy=LoRA2026.05 | 10.5 | 27.1 | 66.2 | |
| LLAMA-MOE v2Architecture Strategy=LLAMA-MOE v22026.05 | 10.2 | 25.5 | 74.1 | |
| Qwen2.5-0.5BModel Scale=0.5B2026.05 | 9.5 | 18.8 | 58.1 | |
| Llama3.2-1BModel Scale=1B2026.05 | 8.1 | 16.2 | 56.5 |