opus-mt-tc-big-sh-en
| Property | Value |
|---|---|
| Parameter Count | 237M |
| License | CC-BY-4.0 |
| Architecture | Transformer-big (Marian NMT) |
| Languages | Serbo-Croatian → English |
| Best BLEU Score | 66.5 (Bosnian-English) |
What is opus-mt-tc-big-sh-en?
opus-mt-tc-big-sh-en is a neural machine translation model developed by Helsinki-NLP as part of the OPUS-MT project. It's specifically designed to translate from Serbo-Croatian languages (including Bosnian, Croatian, and Serbian in both Cyrillic and Latin scripts) to English. The model is built on the transformer-big architecture and trained using data from the OPUS corpus.
Implementation Details
The model utilizes the Marian NMT framework and has been converted to PyTorch using the Hugging Face transformers library. It implements SentencePiece tokenization with 32k vocabulary size and uses FP16 precision for efficient inference.
- Architecture: Transformer-big with 237M parameters
- Training Data: opusTCv20210807+bt dataset
- Tokenization: SentencePiece (spm32k)
- Release Date: 2022-02-25
Core Capabilities
- Multi-variant Translation: Supports Bosnian (Latin), Croatian, Serbian (Cyrillic), and Serbian (Latin) to English
- Strong Performance: BLEU scores ranging from 37.1 to 66.5 across different language pairs
- Production-Ready: Integrated with Hugging Face transformers pipeline for easy deployment
- Batch Processing: Capable of handling multiple sentences simultaneously
Frequently Asked Questions
Q: What makes this model unique?
The model's ability to handle multiple variants of Serbo-Croatian languages with a single architecture makes it particularly valuable for applications requiring robust translation across these closely related languages. Its impressive BLEU scores, particularly for Bosnian-English (66.5), demonstrate its high accuracy.
Q: What are the recommended use cases?
The model is ideal for applications requiring translation from any Serbo-Croatian variant to English, such as content localization, document translation, and cross-lingual information retrieval. It's particularly well-suited for production environments due to its integration with the Hugging Face transformers library.
