m42-health/Llama3-Med42-70B

Med42-v2: A Clinically-Aligned Llama3-70B Model

Med42-v2-70B is a 70 billion parameter large language model developed by M42 Health AI Team, specifically instruction and preference-tuned for medical applications. Built upon the Llama3 architecture, this model aims to expand access to medical knowledge through high-quality generative AI.

Key Capabilities & Performance:

• Superior Medical MCQA Performance: Outperforms GPT-4.0 in most multiple-choice question answering tasks.
• State-of-the-Art MedQA: Achieves a 79.10 zero-shot performance on MedQA, surpassing other openly available medical LLMs.
• Top Clinical Elo Rating: Ranks highest on the Clinical Elo Rating Leaderboard with a score of 1764, significantly outperforming Llama3-70B-Instruct and GPT-4o.
• Instruction-Tuned: Fine-tuned on approximately 1 billion tokens from diverse open-access medical sources, including flashcards, exam questions, and dialogues.
• 8K Context Length: Supports an 8192-token context window for processing medical text.

Intended Use Cases:

• Medical question answering
• Patient record summarization
• Aiding medical diagnosis
• General health Q&A

Important Limitations:

• Not for Clinical Use: The model is not ready for real clinical use and requires extensive human evaluation and safety testing.
• Potential for Harm: May generate incorrect or harmful information and carries a risk of perpetuating biases from training data.

For more details, refer to the research paper.