Translation on Translation (COMET evaluation)
86.31COMET ScoreTeacher (70B SFT)
Evaluation Results
| Method | Links | |
|---|---|---|
| Teacher (70B SFT)Model Size=70B2026.06 | 86.31 | |
| 2-StageModel Size=8B2026.06 | 83.63 | |
| MixModel Size=8B2026.06 | 83.15 | |
| GKDModel Size=8B2026.06 | 83.1 | |
| IGDSModel=Qwen3-8B-Base, Data Selection Strategy=IGDS2026.04 | 83.07 | |
| SFTModel Size=8B2026.06 | 82.87 | |
| FullModel=Qwen3-8B-Base, Data Selection Strategy=Full2026.04 | 82.71 | |
| IFDModel=Qwen3-8B-Base, Data Selection Strategy=IFD2026.04 | 82.69 | |
| 2-StageModel Size=3B2026.06 | 82.67 | |
| LossModel=Qwen3-8B-Base, Data Selection Strategy=Loss2026.04 | 82.66 | |
| RandomModel=Qwen3-8B-Base, Data Selection Strategy=Random2026.04 | 82.61 | |
| ZIPModel=Qwen3-8B-Base, Data Selection Strategy=ZIP2026.04 | 82.57 | |
| MixModel Size=3B2026.06 | 82.41 | |
| OriginalModel=Qwen3-8B-Base, Data Selection Strategy=Original2026.04 | 81.3 | |
| GKDModel Size=3B2026.06 | 79.39 | |
| SFTModel Size=3B2026.06 | 78.91 | |
| 2-StageModel Size=1B2026.06 | 78.65 | |
| IGDSModel=Llama-3.1-8B, Data Selection Strategy=IGDS2026.04 | 78.45 | |
| MixModel Size=1B2026.06 | 78.13 | |
| FullModel=Llama-3.1-8B, Data Selection Strategy=Full2026.04 | 77.68 | |
| IFDModel=Llama-3.1-8B, Data Selection Strategy=IFD2026.04 | 76.82 | |
| ZIPModel=Llama-3.1-8B, Data Selection Strategy=ZIP2026.04 | 74.49 | |
| LossModel=Llama-3.1-8B, Data Selection Strategy=Loss2026.04 | 72.29 | |
| RandomModel=Llama-3.1-8B, Data Selection Strategy=Random2026.04 | 70.78 | |
| GKDModel Size=1B2026.06 | 69.87 | |
| IGDSModel=Gemma-2-2B, Data Selection Strategy=IGDS2026.04 | 68.14 | |
| SFTModel Size=1B2026.06 | 68 | |
| FullModel=Gemma-2-2B, Data Selection Strategy=Full2026.04 | 65.93 | |
| IFDModel=Gemma-2-2B, Data Selection Strategy=IFD2026.04 | 64.19 | |
| ZIPModel=Gemma-2-2B, Data Selection Strategy=ZIP2026.04 | 63.37 | |
| RandomModel=Gemma-2-2B, Data Selection Strategy=Random2026.04 | 62.92 | |
| LossModel=Gemma-2-2B, Data Selection Strategy=Loss2026.04 | 55.31 | |
| OriginalModel=Llama-3.1-8B, Data Selection Strategy=Original2026.04 | 31.11 | |
| OriginalModel=Gemma-2-2B, Data Selection Strategy=Original2026.04 | 26.92 |