alpha-ai/Medical-Diagnosis-COT-Gemma3-270M
alpha-ai/Medical-Diagnosis-COT-Gemma3-270M is a 0.3 billion parameter Gemma 3 model fine-tuned by Alpha AI for medical question answering. It specializes in generating explicit chain-of-thought (CoT) reasoning within tags, followed by a final answer, making it suitable for research into verifiable medical reasoning. The model supports a context length of up to 128K tokens and is optimized for transparent intermediate steps in medical problem-solving.
Loading preview...
Medical-Diagnosis-COT-Gemma3-270M: Chain-of-Thought for Medical Reasoning
This model, developed by Alpha AI, is a fine-tuned version of Google's Gemma 3 (270M parameters) specifically designed for medical question answering. Its core differentiator is the explicit generation of a chain-of-thought (CoT), enclosed within <think>...</think> tags, before providing a final answer. This feature is particularly valuable for understanding the model's reasoning process.
Key Capabilities
- Explicit Chain-of-Thought: Generates detailed reasoning steps, enhancing transparency and interpretability in medical problem-solving.
- Medical Question Answering: Fine-tuned on a dataset of medical reasoning questions, including
FreedomIntelligence/medical-o1-reasoning-SFTand human-annotated data. - Gemma 3 Architecture: Leverages the Gemma 3 base model, supporting a context window of up to 128K tokens.
- Research-Oriented: Ideal for studying CoT interpretability, prompt engineering, and dataset curation in the medical domain.
Good For
- Research on Medical Reasoning: Investigating how LLMs arrive at medical conclusions.
- Internal Tooling: Developing assistants where human review of intermediate reasoning steps is crucial.
- Interpretability Studies: Analyzing the model's thought process for medical diagnoses and treatment planning.
Important Note: This model is a research system and not intended for clinical use or diagnosis/treatment decisions. It may hallucinate facts and does not guarantee adherence to medical guidelines. Users should be aware of potential biases from synthetic training data.