Machine Translation on FLORES-200 ET-EN (test)
17.5BLEUMix-MoE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Mix-MoEArchitecture Strategy=Mix-MoE2026.05 | 17.5 | 38.9 | 67.5 | |
| MOE-LPRArchitecture Strategy=MOE-LPR2026.05 | 17.1 | 37.5 | 66.8 | |
| MoEArchitecture Strategy=MoE2026.05 | 16.9 | 32.9 | 57.5 | |
| Dense PT+SFTTraining Strategy=Single-stage Fine-tuning baseline2026.05 | 16.4 | 37.2 | 67.5 | |
| Full FinetuneTraining Strategy=Full Finetune2026.05 | 15.1 | 36.5 | 67.2 | |
| LLAMA-ProTraining Strategy=LLAMA-Pro2026.05 | 13.8 | 31.2 | 61.5 | |
| LoRATraining Strategy=LoRA2026.05 | 13.2 | 28.5 | 60.1 | |
| LLAMA-MOE v2Architecture Strategy=LLAMA-MOE v22026.05 | 12.8 | 19.8 | 61.2 | |
| Qwen2.5-1.5BModel Scale=1.5B2026.05 | 10.8 | 20.4 | 56.8 | |
| Qwen2.5-0.5BModel Scale=0.5B2026.05 | 8.4 | 15.9 | 51.5 | |
| Llama3.2-1BModel Scale=1B2026.05 | 7.2 | 13.5 | 48.2 |