Machine Translation on FLORES-200
79.23COMETTeacher
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| TeacherBackbone=Gemma-7B-IT2025.09 | 79.23 | — | — | — | |
| AMiDTeacher-Student Architecture=Gemma-7B-It to Gemma-2B-It, alpha=!= +/- 12025.10 | 74.78 | — | — | — | |
| ABKDTeacher-Student Architecture=Gemma-7B-It to Gemma-2B-It2025.10 | 74.21 | — | — | — | |
| TAIDTeacher-Student Architecture=Gemma-7B-It to Gemma-2B-It2025.10 | 74.2 | — | — | — | |
| CSD (T, S)Loss=CSD (T, S)2025.09 | 74.14 | — | — | — | |
| JSLoss=JS2025.09 | 74.05 | — | — | — | |
| JeffreyLoss=Jeffrey2025.09 | 74.02 | — | — | — | |
| KLLoss=KL2025.09 | 73.96 | — | — | — | |
| TVLoss=TV2025.09 | 73.73 | — | — | — | |
| SRKLLoss=SRKL2025.09 | 73.1 | — | — | — | |
| DLD-max (T)Loss=DLD-max (T)2025.09 | 65.28 | — | — | — | |
| SKLLoss=SKL2025.09 | 59.65 | — | — | — | |
| AMiDTeacher-Student Architecture=Qwen2-7B-It to Qwen2-0.5B-It, alpha=!= +/- 12025.10 | 58.31 | — | — | — | |
| TAIDTeacher-Student Architecture=Qwen2-7B-It to Qwen2-0.5B-It2025.10 | 58.17 | — | — | — | |
| ABKDTeacher-Student Architecture=Qwen2-7B-It to Qwen2-0.5B-It2025.10 | 58.07 | — | — | — | |
| DistiLLMTeacher-Student Architecture=Qwen2-7B-It to Qwen2-0.5B-It2025.10 | 57.23 | — | — | — | |
| DistiLLMTeacher-Student Architecture=Gemma-7B-It to Gemma-2B-It2025.10 | 52.83 | — | — | — | |
| RKLLoss=RKL2025.09 | 45.02 | — | — | — | |
| DLD (S)Loss=DLD (S)2025.09 | 21.52 | — | — | — | |
| SmolLM2-1.7bprompting=Zero-shot, data_tokens=11T, gpu_hours=not reported2025.12 | 0.367 | — | — | — | |
| EuroLLM-1.7bprompting=Zero-shot, data_tokens=4T, gpu_hours=500K2025.12 | 0.36 | — | — | — | |
| MiniLingua-1bprompting=Zero-shot, data_tokens=1.5T, gpu_hours=75K2025.12 | 0.343 | — | — | — | |
| Gemma-3-1b-ptprompting=Zero-shot, data_tokens=2T, gpu_hours=∼300K2025.12 | 0.319 | — | — | — | |
| Aya101Model size=13B2026.03 | — | 47.7 | 0.765 | — | |
| BaseBackbone=0.5B2026.05 | — | — | — | 4.685 | |
| BaseBackbone=7B2026.05 | — | — | — | 3.105 | |
| CEBackbone=0.5B, Adapter=LoRA2026.05 | — | — | — | 19.8 | |
| CEBackbone=0.5B, Adapter=SineLoRA2026.05 | — | — | — | 18.4 | |
| CEBackbone=0.5B, Adapter=RandLoRA2026.05 | — | — | — | 14.2 | |
| CEBackbone=0.5B, Adapter=DoRA2026.05 | — | — | — | 19.5 | |
| CEBackbone=7B, Adapter=LoRA2026.05 | — | — | — | 17.4 | |
| CEBackbone=7B, Adapter=SineLoRA2026.05 | — | — | — | 16.8 | |
| CEBackbone=7B, Adapter=RandLoRA2026.05 | — | — | — | 12.5 | |
| CEBackbone=7B, Adapter=DoRA2026.05 | — | — | — | 17.2 | |
| CE + TMKLBackbone=0.5B, Adapter=LoRA2026.05 | — | — | — | 1.5 | |
| CE + TMKLBackbone=0.5B, Adapter=SineLoRA2026.05 | — | — | — | 1.1 | |
| CE + TMKLBackbone=0.5B, Adapter=RandLoRA2026.05 | — | — | — | 0.8 | |
| CE + TMKLBackbone=0.5B, Adapter=DoRA2026.05 | — | — | — | 1.4 | |
| CE + TMKLBackbone=7B, Adapter=LoRA2026.05 | — | — | — | 0.4 | |
| CE + TMKLBackbone=7B, Adapter=SineLoRA2026.05 | — | — | — | 0.6 | |
| CE + TMKLBackbone=7B, Adapter=RandLoRA2026.05 | — | — | — | 0.9 | |
| CE + TMKLBackbone=7B, Adapter=DoRA2026.05 | — | — | — | 0.5 | |
| Gemma3Model size=27B2026.03 | — | 52.4 | 0.778 | — | |
| Llama3.3Model size=70B2026.03 | — | 51 | 0.758 | — | |
| MADLADModel size=10B2026.03 | — | 49.9 | 0.767 | — | |
| NLLBModel size=600M2026.03 | — | 52.4 | 0.799 | — | |
| NLLBModel size=3B2026.03 | — | 55.8 | 0.849 | — | |
| OmniSONAR2026.03 | — | 55.4 | 0.878 | — | |
| SONAR2026.03 | — | 52.9 | 0.817 | — | |
| Tower+Model size=9B2026.03 | — | 45.1 | 0.652 | — |