Fully Open Meditron: First Auditable Pipeline for Clinical LLMs

Researchers have unveiled Fully Open Meditron, marking the debut of a completely open pipeline designed for developing LLM-based clinical decision support systems (CDSS). In contrast to typical 'open' models that only share weights while keeping data provenance and curation methods secret, this pipeline provides full transparency of the entire training process. It features a clinician-reviewed training corpus that integrates eight public medical QA datasets, along with a reproducible framework for data construction and training, and an evaluation protocol aligned with practical use. Additionally, the corpus contains three synthetic extensions vetted by clinicians: exam-style QA, guideline-based QA from 46,469 clinical practice guidelines, and further expansions. This initiative seeks to enhance rigorous and reproducible validation in CDSS, tackling the existing lack of transparency in LLM-based systems.