arasheslamii/biomed-mt-arabic-script

LoRA adapters and merging recipes for biomedical machine translation into low-resource Arabic-script languages (Dari, Urdu) via Arabic and Persian pivots. Models and code for arXiv:2607.22300. # Biomedical MT for Low-Resource Arabic-Script Languages Trained LoRA adapters and merging code for the paper **Biomedical Machine Translation for Low-Resource Arabic-Script Languages via Cross-Lingual Transfer and LoRA Adapter Merging** (arXiv:2607.22300). Abdullah Alabdullah, Arash Eslamighayour, Sarp Harbalioglu, Lifeng Han School of Informatics, University of Edinburgh · Leiden University Medical Center & LIACS, Leiden University The work uses Arabic and Persian as higher-resource pivot languages to improve English→X biomedical machine translation for severely low-resource Arabic-script languages. Domain-specific LoRA adapters are trained on the pivots and then transferred to target languages by three routes: few-shot in-context learning, supervised adaptation on 500 sentences, and zero-data LoRA adapter merging through tensor arithmetic. The central finding is that biomedical domain knowledge can be redirected to closely related languages by weight-space operations alone, with no target-language data at all. This repository is the entry point to the released adapters. It contains no model weights of its own: the adapters live on the Hugging Face Hub and are linked below, and merged adapters are reproduced locally with `merge.py`. --- ## ⚠ Intended use and safety **These are research artifacts. They must not be used for clinical decision-making, patient communication, or any other safety-critical purpose without qualified human review.** Translation errors in healthcare lead to misunderstanding, delayed treatment, and direct patient harm. The adapters are general-purpose instruction-tuned LLMs adapted with LoRA, not systems engineered for translation. They are subject to hallucination, prompt sensitivity, and fluent-but-wrong output. Fluent-but-wrong is the most dangerous failure mode in a medical register, because the output reads as correct to anyone who cannot check it against the source. **Numbers, dosages, units, and dates are not reliably preserve …