|
|
--- |
|
|
license: apache-2.0 |
|
|
tags: |
|
|
- translation |
|
|
- opus-mt |
|
|
- ctranslate2 |
|
|
- quantized |
|
|
language: |
|
|
- multilingual |
|
|
pipeline_tag: translation |
|
|
--- |
|
|
|
|
|
# opus-mt-ja-he-ctranslate2-android |
|
|
|
|
|
This is a quantized INT8 version of `Helsinki-NLP/opus-mt-ja-he` converted to CTranslate2 format for efficient inference. |
|
|
|
|
|
## Model Details |
|
|
|
|
|
- **Original Model**: Helsinki-NLP/opus-mt-ja-he |
|
|
- **Format**: CTranslate2 |
|
|
- **Quantization**: INT8 |
|
|
- **Framework**: OPUS-MT |
|
|
- **Converted by**: Automated conversion pipeline |
|
|
|
|
|
## Files Included |
|
|
|
|
|
- CTranslate2 model files (quantized INT8) |
|
|
- SentencePiece tokenizer files (`source.spm`, `target.spm`) |
|
|
- Integration guide for Android deployment |
|
|
|
|
|
## Usage |
|
|
|
|
|
### With CTranslate2 |
|
|
|
|
|
```python |
|
|
import ctranslate2 |
|
|
import sentencepiece as spm |
|
|
|
|
|
# Load the model |
|
|
translator = ctranslate2.Translator("path/to/model") |
|
|
|
|
|
# Load tokenizers |
|
|
sp_source = spm.SentencePieceProcessor(model_file="source.spm") |
|
|
sp_target = spm.SentencePieceProcessor(model_file="target.spm") |
|
|
|
|
|
# Translate |
|
|
source_tokens = sp_source.encode("Your text here", out_type=str) |
|
|
results = translator.translate_batch([source_tokens]) |
|
|
translation = sp_target.decode(results[0].hypotheses[0]) |
|
|
``` |
|
|
|
|
|
### Android Integration |
|
|
|
|
|
See the included `INTEGRATION_GUIDE.txt` for Android implementation details. |
|
|
|
|
|
## Performance |
|
|
|
|
|
This INT8 quantized version provides: |
|
|
- ~75% reduction in model size |
|
|
- Faster inference speed |
|
|
- Maintained translation quality |
|
|
- Mobile-friendly deployment |
|
|
|
|
|
## Original Model |
|
|
|
|
|
Based on the OPUS-MT project: https://github.com/Helsinki-NLP/Opus-MT |
|
|
|
