Machine Translation on FLORES-200 EN-ZH (test)
24.3BLEUMix-MoE
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| Mix-MoEArchitecture Strategy=Mix-MoE2026.05 | 24.3 | 36.2 | 86.8 | |
| Qwen2.5-1.5BModel Scale=1.5B2026.05 | 23.9 | 34.8 | 85.2 | |
| MOE-LPRArchitecture Strategy=MOE-LPR2026.05 | 22.3 | 33.5 | 86.1 | |
| Qwen2.5-0.5BModel Scale=0.5B2026.05 | 21.6 | 33.5 | 80.5 | |
| MoEArchitecture Strategy=MoE2026.05 | 20.3 | 31.8 | 79.2 | |
| Dense PT+SFTTraining Strategy=Single-stage Fine-tuning baseline2026.05 | 16.4 | 27.5 | 70.5 | |
| Full FinetuneTraining Strategy=Full Finetune2026.05 | 14.8 | 25.5 | 64.5 | |
| LLAMA-ProTraining Strategy=LLAMA-Pro2026.05 | 9.1 | 16.8 | 55.8 | |
| LLAMA-MOE v2Architecture Strategy=LLAMA-MOE v22026.05 | 5.8 | 10.8 | 67.2 | |
| Llama3.2-1BModel Scale=1B2026.05 | 1.4 | 4.2 | 42.1 | |
| LoRATraining Strategy=LoRA2026.05 | 1.1 | 3.5 | 48.5 |