Med42-v2: Clinically-Aligned LLMs

Med42-v2 is a suite of open-access clinical large language models (LLMs) developed by M42 Health, built upon the Llama-3 architecture. This 8 billion parameter version, Llama3-Med42-8B, is instruct and preference-tuned to enhance access to medical knowledge and provide high-quality answers to medical questions. The model has a context length of 8k tokens and processes text-only input to generate text-only output.

Key Capabilities & Performance

• Specialized Medical Knowledge: Fine-tuned on approximately 1 billion tokens from high-quality, open-access medical sources, including flashcards, exam questions, and dialogues.
• Medical Question Answering: Designed to assist with medical inquiries, patient record summarization, and diagnostic aid.
• Competitive Benchmarking: While the 70B version leads, Llama3-Med42-8B shows strong performance in medical evaluations, achieving 62.84 on MedQA and 67.04 on USMLE in zero-shot accuracy. It also has an Elo Score of 924 in the Clinical Elo Rating Leaderboard.

Intended Use Cases

• Medical Question Answering: Providing accurate responses to health-related queries.
• Patient Record Summarization: Assisting in condensing and understanding patient data.
• Aiding Medical Diagnosis: Supporting healthcare professionals in diagnostic processes.
• General Health Q&A: Serving as an AI assistant for general health information.

Limitations & Responsible Use

It is crucial to note that the Med42-v2 suite is not yet ready for real clinical use and requires extensive human evaluation for safety. Users should be aware of the potential for generating incorrect or harmful information and perpetuating biases. The model should not be relied upon for medical decisions or patient care without rigorous safety testing. For more details, refer to the research paper.