EPFLiGHT/Meditron3-Phi4-14B

Meditron3-Phi4-14B: A Specialized Medical LLM

Meditron3-Phi4-14B is a 14.7 billion parameter large language model developed by the OpenMeditron initiative, specifically designed for clinical medicine. It is built on the Phi4 base model and focuses on providing medical insights relevant to diverse settings, including limited-resource and humanitarian contexts.

Key Capabilities & Focus Areas

• Medical Specialization: Trained on a unique mixture of expert-curated medical data, including clinical guidelines, peer-reviewed publications, synthetic differential diagnoses, and medical multiple-choice questions.
• Equitable Representation: Emphasizes contextual diversity and representation of neglected populations and diseases.
• Research Tool: Primarily intended for research to evaluate LLMs in clinical decision-making and information access.
• Foundation Model: Released as a foundation model, it is not instruction-tuned but can be adapted for specific downstream tasks like question-answering through techniques such as in-context learning or fine-tuning.

Performance Highlights

Evaluations on medical multiple-choice question benchmarks show improved performance compared to the base microsoft/phi-4 model:

• MedmcQA: 66.58% (vs. 63.11% for Phi-4)
• MedQA: 69.29% (vs. 62.77% for Phi-4)
• Overall Average: 71.16% (vs. 68.29% for Phi-4)

Important Considerations

• Research-Only: This model is for research purposes only and is not validated for direct medical use or clinical decision-making. It should not be used for self-diagnosis or treatment.
• Static Release: This is a static model trained on an offline dataset, with future enhanced versions planned.