---
language:
  - zh
  - ru
tags:
  - marian
  - opus-mt
  - translation
  - mnn
  - android
  - on-device
license: apache-2.0
library_name: transformers
pipeline_tag: translation
---

# OPUS-MT zh→ru (Custom Trained)

This is a custom-trained Marian-based translation model for zh→ru.
It was fine-tuned from [`Helsinki-NLP/opus-mt-zh-en`](https://huggingface.co/Helsinki-NLP/opus-mt-zh-en)
and converted to MNN format via the [mnn-opus-mt-toolkit](https://github.com/HoSStiA/mnn-opus-mt-toolkit).

## Model Details

| Parameter | Value |
|-----------|-------|
| Architecture | MarianMT (encoder-decoder Transformer) |
| Source language | zh |
| Target language | ru |
| Base model | `Helsinki-NLP/opus-mt-zh-en` |
| BLEU score | 24.55 |
| chrF score | 41.83 |
| Training pairs | 450,000 |
| Training epochs | 25 |

## Usage

### With Hugging Face Transformers
```python
from transformers import AutoModelForSeq2SeqLM, MarianTokenizer

model = AutoModelForSeq2SeqLM.from_pretrained("Hosstia/opus-mt-zh-ru-opus-mp")
tokenizer = MarianTokenizer.from_pretrained("Hosstia/opus-mt-zh-ru-opus-mp")

text = "Your source text here"
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs)
translation = tokenizer.decode(outputs[0], skip_special_tokens=True)
```

### With MNN (on-device Android)
Convert to MNN format using the [mnn-opus-mt-toolkit](https://github.com/HoSStiA/mnn-opus-mt-toolkit):
```bash
./run_all.sh --hf-user <username> --src zh --dst ru --format fp16 \
    --model-repo Hosstia/opus-mt-zh-ru-opus-mp --tokenizer-repo Hosstia/opus-mt-zh-ru-opus-mp
```

## License

Apache License 2.0 (model weights); see the toolkit repository for details.

## Acknowledgments

- [Helsinki-NLP](https://huggingface.co/Helsinki-NLP) for the original OPUS-MT models
- [Marian NMT](https://github.com/marian-nmt/marian) for the training framework
- [Alibaba MNN](https://github.com/alibaba/MNN) for the on-device inference engine
- [Hugging Face](https://huggingface.co) for model hosting and the Transformers library