Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,67 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
language:
|
| 3 |
+
- zh
|
| 4 |
+
- ru
|
| 5 |
+
tags:
|
| 6 |
+
- marian
|
| 7 |
+
- opus-mt
|
| 8 |
+
- translation
|
| 9 |
+
- mnn
|
| 10 |
+
- android
|
| 11 |
+
- on-device
|
| 12 |
+
license: apache-2.0
|
| 13 |
+
library_name: transformers
|
| 14 |
+
pipeline_tag: translation
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# OPUS-MT zh→ru (Custom Trained)
|
| 18 |
+
|
| 19 |
+
This is a custom-trained Marian-based translation model for zh→ru.
|
| 20 |
+
It was fine-tuned from [`Helsinki-NLP/opus-mt-zh-en`](https://huggingface.co/Helsinki-NLP/opus-mt-zh-en)
|
| 21 |
+
and converted to MNN format via the [mnn-opus-mt-toolkit](https://github.com/HoSStiA/mnn-opus-mt-toolkit).
|
| 22 |
+
|
| 23 |
+
## Model Details
|
| 24 |
+
|
| 25 |
+
| Parameter | Value |
|
| 26 |
+
|-----------|-------|
|
| 27 |
+
| Architecture | MarianMT (encoder-decoder Transformer) |
|
| 28 |
+
| Source language | zh |
|
| 29 |
+
| Target language | ru |
|
| 30 |
+
| Base model | `Helsinki-NLP/opus-mt-zh-en` |
|
| 31 |
+
| BLEU score | 24.55 |
|
| 32 |
+
| chrF score | 41.83 |
|
| 33 |
+
| Training pairs | 450,000 |
|
| 34 |
+
| Training epochs | 25 |
|
| 35 |
+
|
| 36 |
+
## Usage
|
| 37 |
+
|
| 38 |
+
### With Hugging Face Transformers
|
| 39 |
+
```python
|
| 40 |
+
from transformers import AutoModelForSeq2SeqLM, MarianTokenizer
|
| 41 |
+
|
| 42 |
+
model = AutoModelForSeq2SeqLM.from_pretrained("Hosstia/opus-mt-zh-ru-opus-mp")
|
| 43 |
+
tokenizer = MarianTokenizer.from_pretrained("Hosstia/opus-mt-zh-ru-opus-mp")
|
| 44 |
+
|
| 45 |
+
text = "Your source text here"
|
| 46 |
+
inputs = tokenizer(text, return_tensors="pt")
|
| 47 |
+
outputs = model.generate(**inputs)
|
| 48 |
+
translation = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
### With MNN (on-device Android)
|
| 52 |
+
Convert to MNN format using the [mnn-opus-mt-toolkit](https://github.com/HoSStiA/mnn-opus-mt-toolkit):
|
| 53 |
+
```bash
|
| 54 |
+
./run_all.sh --hf-user <username> --src zh --dst ru --format fp16 \
|
| 55 |
+
--model-repo Hosstia/opus-mt-zh-ru-opus-mp --tokenizer-repo Hosstia/opus-mt-zh-ru-opus-mp
|
| 56 |
+
```
|
| 57 |
+
|
| 58 |
+
## License
|
| 59 |
+
|
| 60 |
+
Apache License 2.0 (model weights); see the toolkit repository for details.
|
| 61 |
+
|
| 62 |
+
## Acknowledgments
|
| 63 |
+
|
| 64 |
+
- [Helsinki-NLP](https://huggingface.co/Helsinki-NLP) for the original OPUS-MT models
|
| 65 |
+
- [Marian NMT](https://github.com/marian-nmt/marian) for the training framework
|
| 66 |
+
- [Alibaba MNN](https://github.com/alibaba/MNN) for the on-device inference engine
|
| 67 |
+
- [Hugging Face](https://huggingface.co) for model hosting and the Transformers library
|