Model Overview

This model, developed by Matthew Chung, is a 3.1 billion parameter, transformer-based language model fine-tuned from Qwen/Qwen2.5-3B-Instruct. It leverages Generalized Reinforcement Policy Optimization (GRPO) to specialize in medical reasoning tasks, utilizing the Unsloth library for efficient training.

Key Capabilities

• Medical Reasoning: Optimized for understanding and generating responses related to medical reasoning, trained on the FreedomIntelligence/medical-o1-reasoning-SFT dataset.
• GRPO Fine-tuning: Incorporates custom reward functions for semantic correctness and perplexity, enhancing its medical domain performance.
• Educational Use: Primarily intended for educational purposes in medical contexts.

Training Details

The model was trained on 25,117 samples from the FreedomIntelligence/medical-o1-reasoning-SFT dataset over approximately 14 hours on an NVIDIA RTX 3090. Training involved a learning rate of 5e-6, a batch size of 1, and 4-bit quantization, achieving a final loss of 0.001300 and a semantic score of 0.630995.

Limitations and Recommendations

• Educational Use Only: Explicitly stated as not intended for direct medical advice, diagnosis, or treatment recommendations.
• Potential for Inaccuracy: May generate incorrect or misleading medical information.
• Human Oversight Required: Not suitable for high-stakes medical decision-making without professional human oversight.
• Bias Awareness: Users should be aware of potential biases inherited from the training data and verify outputs with medical professionals.