This model is a fine-tuned version of Helsinki-NLP/opus-mt-tc-big-ar-en, adapted for Arabic→English translation as part of the NERDS-NITS submission to the WMT 2026 Low-Resource Arabic–Asian Language Translation Shared Task.
Model Description
- Architecture: Marian Transformer-big (6-layer encoder, 6-layer decoder, hidden size 1024, feed-forward dimension 4096, 16 attention heads per layer, ~240M parameters)
- Tokenization: SentencePiece (inherited from the original OPUS-MT checkpoint, trained on the OPUS parallel corpus)
- Base model:
Helsinki-NLP/opus-mt-tc-big-ar-en - Fine-tuning framework: Hugging Face
Seq2SeqTrainer, minimizing sequence-to-sequence cross-entropy loss - Task: Arabic → English machine translation
- Paper: NERDS-NITS at WMT 2026: Pivot-Based and Direct Neural Machine Translation for Arabic–Asian Low-Resource Language Pairs
Training Data
Fine-tuned on the WMT 2026 Arabic–English parallel corpus.
Evaluation Results
Devtest set
| Direction | BLEU | chrF | COMET |
|---|---|---|---|
| AR→EN | 31.16 | 54.95 | 0.8375 |
Official challenge test set (WMT 2026 organizers)
| Direction | BLEU | chrF2 | TER | COMET | mBERT |
|---|---|---|---|---|---|
| AR→EN | 29.3462 | 55.8251 | 59.8500 | 80.8750 | 81.9188 |
Acknowledgments
We thank the WMT 2026 shared task organizers for providing the datasets and evaluation infrastructure used in this work.