phi_medical_qa_finetune_16bit

A 16-bit merged medical question-answering model fine-tuned with QLoRA on the Medical Meadow MedQA dataset.

• Developed by: A-Kishore
• Base model: unsloth/Phi-3-mini-4k-instruct-bnb-4bit
• License: Apache-2.0
• Task: Medical question answering
• Training approach: QLoRA with PEFT adapters, merged into a 16-bit checkpoint for inference

Model details

This checkpoint was trained to answer medical QA prompts in an instruction-following format. The training workflow uses the unsloth stack together with transformers, peft, trl, bitsandbytes, and datasets.

The dataset used in the notebook is medalpaca/medical_meadow_medqa, and the examples are formatted into a system/user/assistant prompt structure for supervised finetuning.

Intended use

This model is intended for educational and research purposes, and for prototyping medical QA assistants. It should not be used as a substitute for clinical judgment, diagnosis, or treatment recommendations.

Evaluation results

Evaluation was run on 1,018 examples with ROUGE metrics.

Metric	Score
ROUGE-1	0.6212
ROUGE-2	0.5815
ROUGE-L	0.6195

Example usage

from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline

model_id = "A-Kishore/phi_medical_qa_finetune_16bit"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

pipe = pipeline("question-answering", model=model, tokenizer=tokenizer)

prompt = "You are a medical AI. Answer the question clearly and concisely.\n\nQuestion: What is the most likely diagnosis for a patient with fever, rash, and migratory arthritis?"
print(pipe(prompt, max_new_tokens=128, do_sample=False)[0]["generated_text"])

Limitations

The model can produce incorrect or overconfident answers, especially for ambiguous or poorly specified prompts. Review outputs carefully, and do not rely on this model for real-world medical decisions without qualified human oversight.