CAT-Translate: Building Compact Open-Source Models for Japanese-English Translation

The article presents CAT-Translate, a series of compact open-source models for Japanese-English translation, with parameter sizes of 0.8B, 1.4B, 3.3B, and 7B. Utilizing a two-stage supervised fine-tuning approach alongside Multi-Objective GRPO, the models were trained on synthetically generated parallel corpora and evaluated against WMT and real-world benchmarks, showing superior performance in practical applications across various domains compared to larger multilingual models. This development underscores the importance of specialized models for specific language pairs, offering a viable alternative to large-scale multilingual architectures for practitioners focused on targeted translation tasks.