This repository contains a LoRA fine-tuned version of [mistralai/Mistral-7B-Instruct-v0.2] for medical question answering, fine-tuned using QLoRA (4-bit) and PEFT on a custom medical Q&A dataset.

✅ Model Overview

🔬 Base model: mistralai/Mistral-7B-Instruct-v0.2
🧠 Task: Medical question answering
⚙️ Technique: LoRA (Low-Rank Adaptation) + 4-bit QLoRA
📦 Format: Adapter-only (adapter_model.safetensors)

💡 Intended Use

This model is intended for:

Patient education
Clinical assistant prototypes
Biomedical NLP research

⚠️ Not for real-world clinical use. This model is for research/educational purposes only.

🧪 Evaluation

Metric	Before LoRA	After LoRA
BLEU	0.0145	0.0721
F1	0.2457	0.3901

Tested on 100 medical QA samples. Fine-tuning improved answer completeness and accuracy.

🛠️ Training Details Config Value LoRA Rank 16 LoRA Alpha 32 Target Modules q_proj, k_proj, v_proj, o_proj, etc. Epochs 2 Batch Size 8 (effective via gradient accumulation) Max Length 512 tokens Quantization 4-bit (nf4, double quant) Framework Hugging Face Transformers + PEFT

📁 Files Included adapter_model.safetensors – LoRA weights

adapter_config.json – LoRA structure

tokenizer.json, tokenizer_config.json – Tokenizer files

README.md – This file

model_card_data.yaml – Metadata for HF Hub

eval_results.json – Evaluation scores

✍️ Citation

Please credit the original authors of Mistral and cite this fine-tuning work if used in your research or applications.

📌 Developed as part of Phase 1 of a Multimodal Clinical AI Assistant Project.

🙋‍♀️ Author

Dileep Reddy Suram

🚀 Inference Example

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

tokenizer = AutoTokenizer.from_pretrained("your-username/mistral-7b-medical-qa-lora")
base_model = AutoModelForCausalLM.from_pretrained("mistralai/Mistral-7B-Instruct-v0.2", torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(base_model, "your-username/mistral-7b-medical-qa-lora")

def ask_medical_question(question):
    prompt = f"<s>[INST] {question} [/INST]"
    inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
    with torch.no_grad():
        output = model.generate(**inputs, max_new_tokens=150, temperature=0.7)
    return tokenizer.decode(output[0], skip_special_tokens=True).split("[/INST]")[-1].strip()

print(ask_medical_question("What is diabetes?"))

dsuram/mistral-medical-finetuned

🩺 Mistral-7B-Medical-QA-LoRA

✅ Model Overview

💡 Intended Use

🙋‍♀️ Author

🚀 Inference Example

dsuram/
mistral-medical-finetuned