Models
All models are available on Hugging Face:
Evaluation
We conducted evaluation on the translation subsets of the following benchmarks:
We chose these tasks as benchmarks because (1) they are derived from real world applications and (2) are less overoptimized compared to popular datasets (e.g., WMT).
The results are below.
All the models achieved the best scores among all models (including closed source) within their respective sizes for both En-Ja and Ja-En translation tasks.
Table with columns: Model, Avg. BLEU, Avg. BLEU Ja->En, Avg. BLEU En->Ja, BSD (Ja-En), Court (Ja-En), JMed (Ja-En), PFMT (Ja-En), wat-pat-2025 (Ja-En), BSD (En-Ja), JMed (En-Ja), PFMT (En-Ja), wat-pat-2025 (En-Ja)| Model | Avg. BLEU | Avg. BLEU Ja->En | Avg. BLEU En->Ja | BSD (Ja-En) | Court (Ja-En) | JMed (Ja-En) | PFMT (Ja-En) | wat-pat-2025 (Ja-En) | BSD (En-Ja) | JMed (En-Ja) | PFMT (En-Ja) | wat-pat-2025 (En-Ja) |
|---|
| CyberAgent/CAT-Translate-7B | 37.68 | 41.06 | 34.31 | 33.75 | 45.29 |
A detailed experimental evaluation will be present in a technical report.
Usage
The model supports English to Japanese and Japanese to English translation with the following prompt format:
from transformers import pipeline
chat_pipeline = pipeline("text-generation", model="CyberAgent/CAT-Translate-3.3b")
prompt = "Translate the following {src_lang} text into {tgt_lang}.\n\n{src_text}"
src_lang = "Japanese"
tgt_lang = "English"
src_text = "🐈はとてもかわいいの。おててがまるくてふわふわなの。"
user_input = [{"role": "user", "content": prompt.format(src_lang=src_lang, tgt_lang=tgt_lang, src_text=src_text)}]
response = chat_pipeline(user_input, max_new_tokens=512)
print("-" * 20)
print("Source Text:")
print(src_text)
print("Translation:")
print(response[0]['generated_text'][-1]['content'])
Important: You need to apply the chat template to run the model correctly. The template is the same as sarashina2.2-3b-instruct-v0.1.
Why Use Instructions?
Although the model is specialized for machine translation, we require an instruction prompt to invoke the translation capability. This design choice provides better customizability—extending and merging this model is easier this way. Since the model is open source, any extensions are welcome!
Training
We used the sarashina2.2 series (MIT LICENSE) as our pretrained model. While Qwen-3 showed higher benchmark scores, we found that sarashina generated more natural Japanese text that avoided "translationese" patterns. We hypothesized that naturalness is more difficult to learn than translation accuracy, leading us to choose sarashina as our base model.
Our training process involved:
- Synthesizing parallel corpora from monolingual data using large language models
- Two-stage supervised fine-tuning (SFT) approach
- Reinforcement learning with Multi-Objective GRPO (Ichihara et al. 2025)
- LoRA for efficient training
For detailed information about our training methodology, data preparation, and technical specifications, please see TRAINING.md.
License
The model is licensed under the MIT License.
Citation
@misc{jinnai2026cattranslatebuildingcompactopensource,
title={CAT-Translate: Building Compact Open-Source Models for Japanese-English Translation},
author={Yuu Jinnai},
year={2026},
eprint={2606.21413},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2606.21413},
}
Acknowledgments
This project stands on the shoulders of giants. In particular, the following resources significantly helped us develop the model:
