EPFLiGHT/Meditron3-70B

TEXT GENERATIONConcurrent Unit Cost:4Model Size:70BQuant:FP8Context Size:8kTool Calling:SupportedPublished:Jul 12, 2024License:llama3.1Architecture:Transformer0.0K Gated Featherless Exclusive Cold

EPFLiGHT/Meditron3-70B is a 70 billion parameter large language model from EPFLiGHT, specialized in clinical medicine. Built on the Llama-3.1 base, it is trained on expert-curated medical data including clinical guidelines, peer-reviewed publications, and synthetic differential diagnoses. This model is designed for research into enhancing clinical decision-making and access to evidence-based medical information, particularly emphasizing equitable representation and contextual diversity.

Loading preview...

Meditron3-70B: A Clinical LLM for Research

EPFLiGHT/Meditron3-70B is a 70 billion parameter large language model developed by EPFLiGHT, specifically tailored for clinical medicine research. Based on the Llama-3.1 architecture, this model is co-designed with expert clinicians and humanitarian practitioners to ensure equitable representation and contextual diversity, especially for limited-resource settings and neglected populations.

Key Characteristics & Training

  • Specialized Domain: Focused on clinical medicine, aiming to enhance clinical decision-making and access to evidence-based medical information.
  • Comprehensive Training Data: Trained on a unique mixture of expert-curated public data, including:
    • International clinical practice guidelines.
    • Full-text peer-reviewed medical publications.
    • Synthetic conversation data for differential diagnoses.
    • LLM-enhanced medical multiple-choice questions.
  • Research-Only Focus: Intended strictly for research and evaluation purposes; it is not validated for direct medical use or clinical decision-making.

Usage & Adaptability

Meditron3-70B is a foundational model that has not been fine-tuned or instruction-tuned. Developers can adapt it for specific downstream tasks, such as question-answering, using techniques like in-context learning with demonstrations or model fine-tuning with specific datasets. Evaluation is ongoing, with initial assessments using medical multiple-choice questions via lm-harness.