Update README.md

b386247 verified 2 days ago

10.6 kB

	---
	license: mit
	language:
	- ja
	- en
	base_model:
	- sbintuitions/sarashina2.2-0.5b
	---
	# CAT-Translate 🐱

	[![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](https://opensource.org/licenses/MIT)
	[![Hugging Face](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue)](https://huggingface.co/cyberagent/CAT-Translate-0.8b/)

	Tiny Language Model For Japanese and English Bidirectional Translation

	- Purrs on your lap 🐱: Small and efficient! 0.8-7B models that run on edge devices.
	- Swift and Feline Sharp 🐾: Beats TranslateGemma-12B on text-to-text translation quality.
	- Adopt and adapt 🐈: Open source (MIT License) models you can customize and extend.

	<div align="center">
	<img src="CAT-logo.png" alt="Cat sleeping on top of a laptop." width="200">
	</div>

	## Models

	All models are available on Hugging Face:

	- [CAT-Translate-0.8B](https://huggingface.co/cyberagent/CAT-Translate-0.8b/)
	- [CAT-Translate-1.4B](https://huggingface.co/cyberagent/CAT-Translate-1.4b/)
	- [CAT-Translate-3.3B](https://huggingface.co/cyberagent/CAT-Translate-3.3b/)
	- [CAT-Translate-7B](https://huggingface.co/cyberagent/CAT-Translate-7b/)

	## Evaluation

	We conducted evaluation on the translation subsets of the following benchmarks:

	- [The Business Scene Dialogue corpus](https://github.com/tsuruoka-lab/BSD) (BSD)
	- Each conversation is given to the model to translate instead of each sentence.
	- [Court Interpreter](https://github.com/mynlp/court_interpreter) (Court)
	- [JMedBench](https://huggingface.co/datasets/Coldog2333/JMedBench) (JMed)
	- ejmmt subsets are used.
	- [pfmt-bench-fin-ja](https://github.com/pfnet-research/pfmt-bench-fin-ja) (PFMT)
	- [WAT 2025 Patent Translation](https://sites.google.com/view/pat-claims-trans-2025/) (wat-pat-2025)

	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.


	\| 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 \| 30.65 \| 49.86 \| 45.74 \| 16.29 \| 29.62 \| 52.94 \| 38.37 \|
	\| CyberAgent/CAT-Translate-3.3B \| 36.16 \| 37.51 \| 34.80 \| 26.51 \| 42.44 \| 24.47 \| 49.93 \| 44.23 \| 17.21 \| 28.67 \| 53.88 \| 39.44 \|
	\| CyberAgent/CAT-Translate-1.4B \| 33.73 \| 33.26 \| 34.19 \| 31.28 \| 43.84 \| 24.08 \| 36.55 \| 30.57 \| 15.71 \| 26.92 \| 51.53 \| 42.58 \|
	\| Unbabel/Tower-Plus-9B \| 32.41 \| 36.84 \| 27.99 \| 15.43 \| 40.54 \| 29.13 \| 58.00 \| 41.10 \| 10.00 \| 18.80 \| 53.00 \| 30.16 \|
	\| google/translategemma-12b-it \| 32.24 \| 35.81 \| 28.68 \| 31.58 \| 34.30 \| 23.46 \| 48.75 \| 40.97 \| 15.92 \| 21.79 \| 52.53 \| 24.47 \|
	\| CyberAgent/CAT-Translate-3.3B-beta \| 30.60 \| 30.32 \| 30.88 \| 17.20 \| 38.65 \| 23.96 \| 40.58 \| 31.22 \| 16.63 \| 26.68 \| 53.40 \| 26.80 \|
	\| CyberAgent/CAT-Translate-0.8B \| 30.42 \| 29.71 \| 30.68 \| 29.63 \| 33.19 \| 22.96 \| 32.51 \| 30.56 \| 14.60 \| 26.22 \| 50.62 \| 32.87 \|
	\| google/translategemma-4b-it \| 28.09 \| 29.41 \| 26.76 \| 28.86 \| 25.89 \| 21.50 \| 42.65 \| 28.16 \| 14.14 \| 20.68 \| 51.99 \| 20.23 \|
	\| LiquidAI/LFM2.5-1.2B-JP \| 25.47 \| 24.51 \| 26.43 \| 19.06 \| 29.99 \| 22.10 \| 43.61 \| 7.80 \| 14.57 \| 23.85 \| 54.77 \| 12.54 \|
	\| pfnet/plamo-2-translate \| 25.24 \| 25.92 \| 24.57 \| 25.55 \| 28.63 \| 22.90 \| 29.02 \| 23.48 \| 17.35 \| 24.98 \| 32.04 \| 23.89 \|
	\| LiquidAI/LFM2-350M-ENJP-MT \| 24.95 \| 24.91 \| 25.00 \| 10.94 \| 29.56 \| 21.48 \| 41.40 \| 21.17 \| 8.11 \| 22.84 \| 47.53 \| 21.52 \|
	\| mistralai/Ministral-8B-Instruct-2410 \| 24.12 \| 27.52 \| 20.71 \| 19.23 \| 29.21 \| 16.25 \| 50.23 \| 22.69 \| 12.91 \| 16.49 \| 41.66 \| 11.80 \|
	\| nvidia/NVIDIA-Nemotron-Nano-9B-v2-Japanese \| 22.97 \| 22.77 \| 23.18 \| 9.62 \| 34.98 \| 18.01 \| 38.44 \| 12.81 \| 10.62 \| 20.41 \| 42.55 \| 19.13 \|
	\| Rakuten/RakutenAI-2.0-mini-instruct \| 18.43 \| 17.24 \| 19.62 \| 0.11 \| 30.62 \| 18.21 \| 29.34 \| 7.90 \| 5.19 \| 20.36 \| 45.70 \| 7.23 \|
	\| SakanaAI/TinySwallow-1.5B-Instruct \| 15.74 \| 14.99 \| 16.49 \| 4.96 \| 18.93 \| 15.83 \| 26.67 \| 8.58 \| 6.30 \| 17.58 \| 34.07 \| 8.00 \|
	\| llm-jp/llm-jp-3.1-1.8b-instruct4 \| 15.18 \| 16.26 \| 14.11 \| 18.82 \| 2.44 \| 15.67 \| 30.65 \| 13.72 \| 15.38 \| 4.91 \| 25.47 \| 10.65 \|
	\| tencent/HY-MT1.5-1.8B \| 14.49 \| 8.95 \| 20.04 \| 5.50 \| 4.59 \| 4.00 \| 15.67 \| 14.98 \| 6.33 \| 18.13 \| 37.75 \| 17.96 \|
	\| shisa-ai/shisa-v2.1-llama3.2-3b \| 14.27 \| 14.26 \| 14.28 \| 17.08 \| 3.70 \| 8.26 \| 26.86 \| 15.42 \| 13.18 \| 5.54 \| 25.97 \| 12.41 \|
	\| google/gemma-2-2b-jpn-it \| 14.15 \| 16.98 \| 11.32 \| 20.04 \| 8.08 \| 11.27 \| 31.49 \| 14.01 \| 12.37 \| 4.48 \| 16.24 \| 12.21 \|
	\| shisa-ai/shisa-v2.1-lfm2-1.2b \| 13.08 \| 14.02 \| 12.14 \| 20.93 \| 4.95 \| 7.68 \| 26.72 \| 9.80 \| 12.11 \| 5.54 \| 17.60 \| 13.30 \|
	\| microsoft/phi-4 \| 11.92 \| 13.48 \| 10.36 \| 6.10 \| 18.66 \| 2.81 \| 24.86 \| 14.98 \| 3.24 \| 6.97 \| 14.36 \| 16.87 \|
	\| tencent/HY-MT1.5-7B \| 10.56 \| 13.46 \| 7.67 \| 4.99 \| 12.32 \| 5.72 \| 29.53 \| 14.76 \| 0.82 \| 7.80 \| 14.30 \| 7.74 \|
	\| tokyotech-llm/Llama-3.1-Swallow-8B-Instruct-v0.5 \| 10.35 \| 12.42 \| 8.28 \| 24.25 \| 2.30 \| 3.69 \| 14.11 \| 17.74 \| 6.82 \| 2.37 \| 11.21 \| 12.71 \|
	\| Qwen/Qwen2.5-14B-Instruct \| 8.39 \| 9.88 \| 6.89 \| 10.81 \| 4.70 \| 4.27 \| 11.18 \| 18.46 \| 4.01 \| 3.69 \| 13.42 \| 6.42 \|
	\| meta-llama/Llama-3.2-3B-Instruct \| 6.06 \| 9.90 \| 2.23 \| 18.60 \| 0.41 \| 2.72 \| 16.62 \| 11.17 \| 1.44 \| 1.10 \| 4.50 \| 1.87 \|

	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:

	```python
	from transformers import pipeline

	# Load the model
	chat_pipeline = pipeline("text-generation", model="CyberAgent/CAT-Translate-0.8b")

	# Define the prompt template
	prompt = "Translate the following {src_lang} text into {tgt_lang}.\n\n{src_text}"

	# Example: Japanese to English
	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 [sarashina2.2-0.5b-instruct-v0.1](https://huggingface.co/sbintuitions/sarashina2.2-0.5b-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](https://huggingface.co/collections/sbintuitions/sarashina22) ([MIT LICENSE](https://huggingface.co/sbintuitions/sarashina2.2-0.5b/blob/main/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)](https://arxiv.org/abs/2509.22047)
	- LoRA for efficient training

	For detailed information about our training methodology, data preparation, and technical specifications, please see [TRAINING.md](TRAINING.md).

	## License

	The model is licensed under the [MIT License](LICENSE).

	## Citation

	```bibtex
	@misc{cat-translate-2026,
	title={CAT-Translate: Tiny Language Model For Japanese and English Bidirectional Translation},
	author={Yuu Jinnai},
	year={2026},
	url={https://huggingface.co/collections/cyberagent/cat-translate}
	}
	```

	## Acknowledgments

	This project stands on the shoulders of giants. In particular, the following resources significantly helped us develop the model:

	- [sarashina](https://huggingface.co/sbintuitions) by SB Intuitions
	- [gpt-oss](https://huggingface.co/openai/gpt-oss-20b) by OpenAI
	- [MetricX](https://huggingface.co/google/metricx-24-hybrid-xl-v2p6-bfloat16) by Juraj Juraska et al.
	- [Duplodocus](https://github.com/allenai/duplodocus) by AllenAI
	- [fastText](https://github.com/facebookresearch/fastText) by Facebook Research
	- [COMET](https://huggingface.co/Unbabel/wmt22-comet-da) by Ricardo Rei et al.
	- [sacrebleu](https://github.com/mjpost/sacrebleu) by Matt Post
	- Mitsuki Sakamoto for deploying the model with UI for internal testing