| Property | Value |
|---|---|
| License | MIT |
| Base Model | facebook/mbart-large-cc25 |
| Training Data | JESC Dataset |
| BLEU Score | 18.18 |
What is mbart-ja-en?
mbart-ja-en is a specialized Japanese-to-English translation model built upon Facebook's mBART architecture. This model represents a significant advancement in neural machine translation, specifically optimized for Japanese to English language pairs through fine-tuning on the JESC (Japanese-English Subtitle Corpus) dataset.
Implementation Details
The model utilizes the Transformers library and employs a SentencePiece tokenizer specifically trained on the JESC dataset. It builds upon the mBART-large-cc25 architecture, which has been proven effective for multilingual translation tasks.
- Leverages the powerful mBART architecture for sequence-to-sequence translation
- Implements custom SentencePiece tokenization for optimal Japanese text processing
- Achieves a BLEU score of 18.18 on the Kyoto University JEC Basic Sentence Data
Core Capabilities
- High-quality Japanese to English translation
- Support for batch processing and early stopping
- Efficient tokenization optimized for Japanese text
- Maximum output length of 48 tokens
Frequently Asked Questions
Q: What makes this model unique?
This model stands out due to its specialized fine-tuning on the JESC dataset, making it particularly effective for Japanese-to-English translation tasks. The use of custom SentencePiece tokenization further enhances its ability to handle Japanese text accurately.
Q: What are the recommended use cases?
The model is ideal for Japanese-to-English translation tasks, particularly in scenarios requiring accurate translation of general text. It's especially suitable for applications in subtitle translation, given its training on the JESC dataset.