| Property | Value |
|---|---|
| Parameter Count | 238M |
| Model Type | Translation (transformer-big) |
| License | CC-BY-4.0 |
| Language Support | Danish, Swedish → Arabic |
| Framework | PyTorch/Transformers |
What is opus-mt-tc-big-gmq-ar?
opus-mt-tc-big-gmq-ar is a specialized neural machine translation model developed by the Helsinki-NLP group for translating North Germanic languages (specifically Danish and Swedish) to Arabic. Built on the MarianNMT framework and converted to PyTorch, this model achieves impressive BLEU scores ranging from 16.8 to 19.9 on the FLORES101 test set.
Implementation Details
The model utilizes a transformer-big architecture and employs SentencePiece tokenization with 32k vocabulary size. It requires specific language tokens (e.g., >>ara<<) at the beginning of input sentences to indicate the target Arabic dialect.
- Supports multiple Arabic variants including Modern Standard Arabic (ara), Egyptian Arabic (arz), and Levantine Arabic (apc)
- Trained on the opusTCv20210807 dataset
- Implements FP16 precision for efficient inference
Core Capabilities
- High-quality translation from Danish to Arabic (BLEU: 19.9, chrF: 0.528)
- Swedish to Arabic translation (BLEU: 19.3, chrF: 0.519)
- Support for multiple Arabic dialects
- Efficient processing with transformer architecture
Frequently Asked Questions
Q: What makes this model unique?
This model specializes in North Germanic to Arabic translation, supporting multiple source languages and target Arabic dialects with state-of-the-art performance. It's part of the larger OPUS-MT project, making professional-grade translation accessible for less-common language pairs.
Q: What are the recommended use cases?
The model is ideal for translating Danish and Swedish content to Arabic, particularly useful for document translation, content localization, and cross-cultural communication. It's specifically designed for production environments requiring high-quality Arabic translations.