Fully Open Meditron: An Auditable Pipeline for Clinical LLMs

Researchers introduce Fully Open Meditron, the first completely transparent pipeline for building clinical AI systems that exposes training data, curation procedures, and generation methods. The framework achieves state-of-the-art performance on medical benchmarks while maintaining full auditability and reproducibility, addressing a critical gap in transparent healthcare AI.

The introduction of Fully Open Meditron represents a significant shift in how clinical AI systems are developed and validated. Unlike existing 'open' models that release only weights while hiding crucial training details, this pipeline achieves transparency across the entire stack—from data sourcing through evaluation. This matters because clinical decision support systems carry profound responsibility: opaque models cannot be properly audited for bias, errors, or unsafe recommendations that could harm patients.

The healthcare industry has struggled with the tension between model performance and transparency. Proprietary systems offer optimization but lack scrutiny, while truly open approaches have historically lagged in capability. Meditron bridges this gap by establishing rigorous standards: clinician-vetted training data, systematic decontamination protocols, and evaluation by medical professionals rather than purely algorithmic metrics. The integration of 46,469 clinical practice guidelines and multi-physician validation demonstrates that domain expertise can enhance both safety and accuracy.

For healthcare institutions and regulators, this work signals that transparent, auditable AI systems need not compromise on performance. The documented improvements across multiple base models (Apertus-70B achieving +6.6 point gains, Gemma-3 outperforming MedGemma) prove that the framework is generalizable. This establishes a template that other institutions and researchers can replicate, potentially accelerating adoption of transparent clinical AI.

The precedent carries broader implications: it demonstrates that reproducible pipelines with institutional oversight can achieve competitive results in specialized domains. Future clinical AI development may increasingly require similar transparency standards, particularly as regulatory bodies examine AI safety in healthcare.

• →Fully Open Meditron is the first clinical AI pipeline with complete transparency across data, training, and evaluation stages.
• →The framework uses clinician-audited training data, systematic decontamination, and four-physician validation to ensure auditability without sacrificing performance.
• →Multiple base models improved significantly with Meditron fine-tuning, with Apertus-70B reaching 53.8% on medical benchmarks versus 47.2% baseline.
• →Integration of 46,469 clinical practice guidelines and synthetic clinical vignettes demonstrates how domain expertise improves both safety and capability.
• →The reproducible pipeline establishes a replicable template for transparent healthcare AI that regulators and institutions can adopt as a standard.