drwlf / Medra

Overview

Medra is a purpose-built, lightweight medical language model designed to assist in clinical reasoning, education, and dialogue modeling.
Built on top of Gemma 3 , Medra is the first step in a long-term project to create deployable, interpretable, and ethically aligned AI support systems for medicine.

It is compact enough to run on consumer hardware.
Capable enough to support nuanced medical prompts.
And principled enough to never pretend to replace human judgment.

Medra is not a chatbot.
It is a cognitive tool —a reasoning companion for students, clinicians, and researchers exploring how AI can help illuminate the complexity of care without oversimplifying it.

Purpose & Philosophy

Medra was developed to fill a crucial gap in the current AI landscape:

While many general-purpose LLMs excel at open-domain conversation, very few are optimized for structured, medically relevant reasoning.
Even fewer can run locally , offline, and in real-time—particularly in environments where access to massive models is impractical or unethical.

Medra aims to provide:

• Interpretable outputs for case simulation and review
• Support for differential diagnosis exploration
• A reflective partner for medical students
• A framework for reasoning refinement in applied clinical contexts

This project is rooted in the belief that AI in healthcare must be transparent , educational , and augmentative —not autonomous, extractive, or misleading.

Key Capabilities

• Lightweight Clinical Reasoning Core
  Medra is fine-tuned to support structured medical queries, diagnostic steps, SOAP formatting, and clinical questioning strategies.

• Local and Mobile Friendly
  Offered in GGUF (Q4, Q8, BF16), Medra can run on local devices via Ollama, LM Studio, KoboldCpp, and other local inference engines—no API needed.

• Data & Alignment
  Trained on medical content including PubMed-derived literature, reasoning datasets (e.g. R1 distilled), clinical notes, and prompt structures modeled after real-world physician interactions.

• High Interpretability
  Designed for transparency and reflection—not black-box decision-making. Medra works best when prompted like a partner, not a prophet.

• Designed for Ethical Integration
  Built with the explicit goal of remaining aligned, cautious, and useful for human-in-the-loop medical settings.

Intended Use

• Medical education and exam-style reasoning
• Case-based learning simulation
• AI health assistant prototyping
• Dialogue modeling in therapeutic or diagnostic contexts
• As a tool for thinking alongside , not thinking instead of

Limitations

• Medra is not a licensed medical professional.
  It is not intended for real-world diagnosis, treatment planning, or patient interaction without human oversight.

• The model may hallucinate, oversimplify, or present outdated medical knowledge in edge cases.

• Medra is not currently equipped with long-term memory, real-world clinical data access, or the authority to guide care.

• It is a prototype. A foundation. Not a finished replacement for expertise.

Technical Details

• Base model: Gemma 3
• Fine-tuning stages: Instructional tuning (STF); RLHF planned in upcoming release
• Data domains: Medical Q&A, differential diagnosis formats, clinical conversation datasets, PubMed-derived material
• Supported inference engines: Ollama, LM Studio, KoboldCpp, GGML-compatible platforms
• Quantization formats: Q4, Q8, BF16

License

Apache 2.0

The Medra Family

Medra is part of a growing family of medical reasoning models:

• Medra — Gemma-based compact model for lightweight local inference
• MedraQ — Qwen 3-based, multilingual and adaptive version
• MedraOmni — Future flagship model built on Qwen 2.5 Omni with full multimodal support

Each model in the series is purpose-built, ethically scoped, and focused on responsible augmentation of healthcare knowledge—not its replacement.

Final Note

Medra exists because medicine deserves tools that reflect care , not just computation.
It is small, but intentional.
Experimental, but serious.
And it was built with one purpose:

To make intelligent care more accessible, more transparent, and more aligned with the human beings it’s meant to serve.

Uploaded finetuned model

• Developed by: drwlf
• License: apache-2.0
• Finetuned from model : unsloth/gemma-3-4b-it-unsloth-bnb-4bit

This gemma3 model was trained 2x faster with Unsloth and Huggingface's TRL library.