mishbahul-hub/WMT-26

Machine translation on low-resource language pairs, built for the WMT 2026 WMT'26 Low-Resource Neural Machine Translation Getting Started  |  Structure  |  Models  |  Usage  |  Results  |  License # WMT-26 **WMT-26** is a research codebase for **machine translation on low-resource language pairs**, built for the WMT 2026 (Conference on Machine Translation) shared-task setting. It benchmarks and fine-tunes modern open multilingual models across two directions in each pair: 1. **Arabic ↔ English** 2. **Arabic ↔ Hindi** The repository is organized as a set of reproducible Jupyter notebooks (one per model/direction) plus a small, modular Python package for **parameter-efficient fine-tuning (LoRA)**. Every experiment is scored with the three metrics standard in MT evaluation today — **BLEU**, **chrF++**, and **COMET** — so zero-shot baselines and fine-tuned checkpoints can be compared on equal footing. > **Why this matters.** Arabic–Hindi in particular is a genuinely low-resource direction with little high-quality parallel data. This repo treats it as a first-class problem: it establishes strong zero-shot baselines from large pretrained models, then measures how much LoRA fine-tuning on a modest parallel corpus closes the gap. --- ## 📑 Table of Contents - Overview - Repository Structure - Models - Evaluation Metrics - Getting Started - Usage - Results - Roadmap - Contributing - License - Acknowledgements --- ## 🔭 Overview The project is split by **language pair**, and within each pair by **approach**: | Language pair | Approach | Backbone family | Notebooks / code | | :--- | :--- | :--- | :--- | | Arabic ↔ English | Pretrained inference | Helsinki-NLP **OPUS-MT** | `Arabic - English/` | | Arabic ↔ Hindi | **Zero-shot** | Meta **NLLB-200**, Google **MADLAD-400** | `Arabic - Hindi/Zero-shot/` | | Arabic ↔ Hindi | **Fine-tuned** | **NLLB-200** (600M, 1.3B + LoRA) | `Arabic - Hindi/Fine-tuned/` | | Arabic → Hi …