Machine Translation on FLORES-200 EN-FI (test)
9.9BLEUMix-MoE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Mix-MoEArchitecture Strategy=Mix-MoE2026.05 | 9.9 | 19.2 | 56.8 | |
| Qwen2.5-1.5BModel Scale=1.5B2026.05 | 8.9 | 12.5 | 48.5 | |
| MOE-LPRArchitecture Strategy=MOE-LPR2026.05 | 7.7 | 15.5 | 55.2 | |
| MoEArchitecture Strategy=MoE2026.05 | 7.4 | 14.2 | 52.1 | |
| Llama3.2-1BModel Scale=1B2026.05 | 7.2 | 11.5 | 50.1 | |
| Qwen2.5-0.5BModel Scale=0.5B2026.05 | 6.4 | 9.8 | 43.2 | |
| Dense PT+SFTTraining Strategy=Single-stage Fine-tuning baseline2026.05 | 5.7 | 9.5 | 46.5 | |
| Full FinetuneTraining Strategy=Full Finetune2026.05 | 3.8 | 7.2 | 43.1 | |
| LLAMA-ProTraining Strategy=LLAMA-Pro2026.05 | 3.5 | 6.2 | 43.5 | |
| LLAMA-MOE v2Architecture Strategy=LLAMA-MOE v22026.05 | 3.5 | 5.8 | 47.1 | |
| LoRATraining Strategy=LoRA2026.05 | 3.3 | 5.5 | 47.5 |