Machine Translation (DE→EN) on IWSLT 2017 (test)
37.21BLEUUniR
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| UniRBackbone=Llama3.2-3B, Trained Model=1B2025.05 | 37.21 | 82.82 | 95.89 | 71.97 | |
| GPT4.1-nanoTrained Model=-2025.05 | 37.14 | 83.7 | 97 | 72.61 | |
| GRPO FullBackbone=Llama3.2-3B, Trained Model=3B2025.05 | 35.52 | 80.43 | 92.35 | 69.43 | |
| Google translatorTrained Model=-2025.05 | 32.95 | 81 | 92.9 | 68.95 | |
| Qwen2.5-3B-InstructTrained Model=-2025.05 | 32.75 | 82.66 | 95.7 | 70.37 | |
| Backbone + 1BBackbone=Llama3.2-3B, Trained Model=-2025.05 | 31.82 | 81.75 | 94.74 | 69.44 | |
| GRPO LoRABackbone=Llama3.2-3B, Trained Model=3B2025.05 | 31.39 | 81.73 | 94.95 | 69.36 | |
| Backbone onlyBackbone=Llama3.2-3B, Trained Model=-2025.05 | 30.65 | 81.27 | 94.53 | 68.82 | |
| alpha-entmaxactivation=alpha-entmax2019.08 | 29.9 | — | — | — | |
| 1.5-entmaxactivation=1.5-entmax2019.08 | 29.83 | — | — | — | |
| softmaxactivation=softmax2019.08 | 29.79 | — | — | — | |
| LABOModel=Transformer2023.05 | 28.4 | — | — | — | |
| LSModel=Transformer2023.05 | 28 | — | — | — | |
| TransformerModel=Transformer2023.05 | 26.9 | — | — | — |