EPFLiGHT/Meditron3-Gemma2-2B

Meditron3-Gemma2-2B: A Medical LLM for Research

EPFLiGHT/Meditron3-Gemma2-2B is a 2.6 billion parameter large language model developed by the OpenMeditron initiative, specifically designed for clinical medicine. Fine-tuned from Google's Gemma2-2B, this model emphasizes equitable representation, contextual diversity, and real-world evidence-based guidelines, with a particular focus on limited-resource and humanitarian settings.

Key Capabilities & Features

• Medical Specialization: Trained on a unique mixture of expert-curated data including clinical guidelines, peer-reviewed medical publications, synthetic differential diagnoses, and LLM-enhanced medical MCQs.
• Research Focus: Intended for research to evaluate the potential of LLMs in clinical decision-making and access to medical information.
• Adaptability: As a foundation model, it can be adapted for specific downstream tasks like Q&A through in-context learning or further fine-tuning.
• Contextual Diversity: Training data aims to represent neglected populations and diseases, moving beyond standard medical contexts.

Performance & Evaluation

Evaluated on medical multiple-choice questions, Meditron3-Gemma2-2B shows competitive performance against google/gemma-2-2b-it on benchmarks like MedmcQA, MedQA, and PubmedQA, with an average slight improvement. The developers note that while MCQs are valuable, they are developing platforms for expert feedback to assess real-world utility, empathy, and alignment to local guidelines.

Important Considerations

• Research-Only Model: This model is not validated for medical use and should not be used for clinical decision-making, diagnosis, or treatment without appropriate validation and regulatory approval.
• Static Model: This is a static model trained on an offline dataset; future enhanced versions are planned.