TibbScholar is a specialized, 3-billion-parameter language model fine-tuned for medical question-answering. It is based on Meta's Llama-3.2-3B model and is designed to serve as an informational tool for educational and research purposes.
This model was trained on a 100,000-record subset of the MIRIAD-4.4M dataset to produce concise, structured answers to medical questions.
Safety and Evaluation Status
TibbScholar is an educational and research prototype—not a source of medical advice, diagnosis, or treatment. Its answers can be incomplete, outdated, or wrong and require verification by qualified clinicians and authoritative medical sources. I have not established clinical reliability, and the model should not be used for patient-care decisions or emergencies.
The project demonstrates a fine-tuning workflow. A documented clinical benchmark, expert review, bias analysis, and safety evaluation have not yet been completed, so this page does not claim medical accuracy or dependable real-world use.
Training Setup
-
Base Model: unsloth/Llama-3.2-3B-unsloth-bnb-4bit
-
Fine-tuning Dataset: A 100,000 record subset of MIRIAD-4.4M.
-
Prompt Format: A simple Question/Answer structure (see below).
-
Training Framework: Fine-tuned using Unsloth AI's library with QLoRA for efficient training.
QLoRA made it possible to adapt the quantized base model while training a much smaller set of parameters than full-model fine-tuning. The project focuses on the reproducible model-adaptation workflow rather than presenting the result as a clinical system.
Intended Use
TibbScholar is suitable for inspecting an educational fine-tuning experiment, comparing structured response behavior, and studying a compact open model artifact. It is not intended to determine a diagnosis, recommend treatment, triage an emergency, or replace a qualified clinician.
Inspect the Model
The public Hugging Face page provides the model artifact and its base-model context. The live link on this case study leads directly to that inspectable project output.
