MohamedAhmedAE/Llama-3.1-8B-Instruct-Medical-Finetuned-merged

Overview

This model, Med-LLaMA3.1-8B, is an 8 billion parameter Llama 3.1-based instruction-tuned language model developed by Mohamed Ahmed Abo El-Enen, Sally S. Ismail, and Taymoor Mohamed Nazmy. It is part of the Med-LLaMA3 family, adapted for the medical domain through QLoRA fine-tuning, with the LoRA weights merged directly into the base model. This allows for direct loading with transformers without additional steps. The model features a 128K token context window and was trained on a curated medical instruction dataset of over 1.5 million samples.

Key Capabilities

• Medical Question Answering: Excels in both multiple-choice and open-ended medical questions, achieving a mean accuracy of 75.71% on MMLU medical subsets.
• Clinical Knowledge & Decision Support: Designed for clinical knowledge lookup and assisting with clinical decision support.
• Clinical Named-Entity Recognition: Demonstrates high precision and recall for extracting diseases and procedures from clinical texts.
• Efficiency: Optimized for consumer-grade hardware, requiring approximately 5.2 GB GPU memory for 4-bit inference, with speeds around 30 tokens/s.

Good For

• Medical question answering and clinical knowledge lookup.
• Clinical decision support assistance (with human review).
• Clinical named-entity recognition (Disease / Procedure extraction).
• Research baselines for parameter-efficient fine-tuning of LLaMA models in healthcare.

Limitations

It is crucial to note that this model is a research artifact and not a medical device. It should not be used for autonomous diagnosis, treatment, or direct patient care. The model can hallucinate and over-elaborate, and its performance is comparable, not superior, to institutionally trained models like LLaMA3-Med42-8B. Human verification is essential due to potential critical errors.