This repository contains a PEFT LoRA adapter fine-tuned (QLoRA-style) on the MedQuAD Medical QnA dataset to answer medical questions using a structured prompt format.

⚠️ Medical disclaimer: This model is for educational/research use only and does not provide medical advice. Outputs may be incomplete or incorrect. For medical concerns, consult a licensed professional.

Model Details

Model Description

Adapter type: LoRA (trained with a QLoRA setup: quantized base model + LoRA adapters)
Base model: BioMistral/BioMistral-7B
Task: Medical question answering / text-generation (instruction-style)
Language: English
Training framework: Hugging Face Transformers + PEFT + bitsandbytes
LoRA configuration (core):
- Rank (r): 8
- Alpha: 16
- (Dropout depends on the training notebook configuration if set.)

Developed by

Developed by: Aravind Gopisetty
Model type: Causal Language Model adapter (PEFT LoRA)

Model Sources

Base model: BioMistral/BioMistral-7B
Dataset: keivalya/MedQuad-MedicalQnADataset
Training / code: https://github.com/aravindgopisetty/biomistral-medqa-qlora

What’s in this repo?

This is not a full model upload. It contains only the LoRA adapter weights (a lightweight “delta” on top of the base model).
To use it, you load the base model and then attach this adapter.

Intended Uses

Direct Use

Generate medical QA style answers for educational and research experimentation.
Demonstrate QLoRA/PEFT fine-tuning workflow on a biomedical LLM under limited GPU constraints.

Out-of-Scope Use

Medical diagnosis, treatment decisions, emergency guidance, or any high-stakes medical use.
Use in clinical settings without robust safety evaluation and expert oversight.

Bias, Risks, and Limitations

The model may hallucinate or produce outdated/incorrect medical information.
Dataset answers may vary in style/length; ROUGE may not fully capture factual correctness.
This adapter inherits limitations and potential biases of the base model and training data.

Recommendation: Add safety guardrails (refusal policies, medical disclaimers, retrieval augmentation) for any downstream demo.

How to Get Started

1) Install dependencies

bash

pip install -U transformers peft bitsandbytes accelerate
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

base_model_id = "BioMistral/BioMistral-7B"
adapter_id    = "aravindgopisetty123/biomistral-medqa-qlora-adapter"

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_use_double_quant=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.float16,
)

tokenizer = AutoTokenizer.from_pretrained(base_model_id, use_fast=True)
tokenizer.pad_token = tokenizer.eos_token

base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    quantization_config=bnb_config,
    device_map="auto",
    use_safetensors=False,
)

model = PeftModel.from_pretrained(base_model, adapter_id)
model.eval()

prompt = """From the MedQuad MedicalQA Dataset: Given the following medical question and question type, provide an accurate answer:

### Question type:
symptoms

### Question:
What are the symptoms of Norrie disease ?

### Answer:
"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    out = model.generate(
        **inputs,
        max_new_tokens=200,
        do_sample=False,
        temperature=0.0,
        pad_token_id=tokenizer.eos_token_id,
        eos_token_id=tokenizer.eos_token_id,
        repetition_penalty=1.1,
    )

print(tokenizer.decode(out[0], skip_special_tokens=True))


Training Details
Training Data

Dataset: keivalya/MedQuad-MedicalQnADataset

Dataset fields used: qtype, Question, Answer

Prompt format:

Question type + question → model generates answer

Training Procedure

Approach: QLoRA-style adapter fine-tuning

Base model loaded in 4-bit NF4 via bitsandbytes

LoRA adapter trained using PEFT

Optimizer: paged AdamW 8-bit (bitsandbytes), gradient checkpointing enabled (for memory efficiency)

Compute / Environment (edit if you want)

Platform: Kaggle (free GPU)

GPU: <T4 or P100>

Precision: fp16/bf16 depending on runtime availability

Evaluation
Metric

ROUGE-L (text overlap) computed on a held-out test split.

Results

Evaluated on 165 held-out samples:

ROUGE-L mean: 0.2809

ROUGE-L median: 0.2387

Note: ROUGE can underestimate performance in free-form QA since correct answers may be phrased differently than references.

Citation

If you use this adapter, please cite the base model and dataset:

BioMistral/BioMistral-7B

keivalya/MedQuad-MedicalQnADataset