Model Overview

This model, ArindamSingh/gemma-3-1b-it-medical-o1-reasoning-finetune-16bit, is a 1 billion parameter Gemma-3 1B-IT variant developed by ArindamSingh. It has been specifically fine-tuned using the FreedomIntelligence Medical-O1 Reasoning (20k Q-CoT-A) dataset to enhance its capabilities in clinical and biomedical reasoning. The fine-tuning process utilized Unsloth for accelerated LoRA training, resulting in merged weights that require no adapter step, making it a direct replacement for the base Gemma model.

Key Capabilities

• Chain-of-Thought (CoT) Answers: Generates step-by-step rationales for medical queries, improving reasoning transparency.
• Compact Footprint: With merged 16-bit weights, the model is less than 2 GB and can run efficiently with minimal VRAM (≤ 4 GB).
• Enhanced Medical QA: Shows improved performance on medical question-answering tasks, such as a 57% QA-F1 score on PubMedQA compared to 46% for the base Gemma-1B-IT.

Intended Use Cases

• Medical Education: Assisting in learning and understanding complex medical concepts.
• PubMed Triage: Aiding in the initial categorization or filtering of medical literature.
• Coding Suggestions: Generating suggestions for medical codes like ICD or SNOMED.
• Research Ideation: Supporting the brainstorming phase of medical research.

Limitations

• Potential for hallucination of citations or outdated guidelines.
• Primarily English-centric, with limited coverage for pediatrics or rare diseases.
• Not a medical device: Should not be used for automated diagnosis, emergency triage, or any decision requiring regulatory clearance; always requires human review.