From Uncharted Data to Edge Adaptation: Fine-Tuning Gemma 3 270M for Multilingual Indic RxNorm Medicine Normalization

This paper presents IndicRxNorm-LexMap-15K, a 15,000-row multilingual Indic medicine terminology instruction dataset grounded in RxNorm/RxNav metadata, covering Hindi, Bengali, Hinglish, and Banglish across six task families: medicine NER, RxNorm normalization, RxCUI entity linking, drug-field extraction, terminology summaries, and safety-boundary refusal. Using Adaption Labs' Adaptive Data platform, the dataset was refined from quality grade B to A. We fine-tuned Gemma 3 270M with LoRA/PEFT using Unsloth, improving JSON parse rate from 7.22% to 71.67% and RxCUI exact-match rate from 2.33% to 74.42% on a held-out sample. We also release a Kaggle Community Benchmark for structured medicine normalization evaluation. Dataset, LoRA adapter, and benchmark are publicly available. This work is for research and terminology normalization only — not diagnosis, prescription, or treatment advice.