Meditron
Technical Summary
Meditron is an open-source suite of large language models (7B and 70B parameters) specifically tailored for the medical domain. Developed by EPFL, Meditron is adapted from Llama-2 to provide clinical reasoning capabilities.
Key Capabilities
- Curated Medical Corpus: Pre-trained on a comprehensive corpus of high-quality medical data, including PubMed abstracts, full-text articles from PMC, and curated clinical guidelines from various medical sources.
- Clinical Reasoning: Demonstrates strong performance on standard medical QA benchmarks like MedQA, MedMCQA, and PubMedQA, approaching or exceeding the performance of proprietary models in some tasks.
- Accessible Scaling: Offered in a 7B parameter version for localized, low-resource deployment on single GPUs, and a 70B parameter version for enterprise-grade clinical decision support.
Usage in Healthcare
Meditron allows healthcare institutions to deploy a powerful, medically-tuned LLM entirely on-premise without sending sensitive Protected Health Information (PHI) to third-party APIs. This makes it an ideal foundational model for developing compliant AI assistants, automated scribes, and diagnostic decision-support tools inside hospital firewalls.
Model Card Details
Architecture
Transformer-based Large Language Model (7B/70B variants) adapted from Llama-2.
Intended Use Cases
Clinical reasoning, medical QA, and diagnostic decision support.
Training Data
Pre-trained on PubMed abstracts, full-text articles from PMC, and curated clinical guidelines.