Adilbai/medical-qa-t5-lora

🏥 Medical QA T5 LoRA Model

A fine-tuned T5 model with LoRA for medical question-answering tasks

🚀 Quick Start • 📊 Performance • 💻 Usage • 🔬 Evaluation

---

📋 Model Overview

This model is a fine-tuned version of Google's T5 (Text-to-Text Transfer Transformer) optimized for medical question-answering tasks using Low-Rank Adaptation (LoRA) technique. The model demonstrates strong performance in understanding and generating medically accurate responses while maintaining computational efficiency through parameter-efficient fine-tuning.

🎯 Key Features

• 📚 Medical Domain Expertise: Fine-tuned specifically for healthcare and medical contexts
• ⚡ Efficient Training: Uses LoRA for parameter-efficient fine-tuning
• 🎯 High Accuracy: Achieves strong performance across multiple evaluation metrics
• 🔄 Versatile: Handles various medical question types and formats

---

🚀 Quick Start

Installation

pip install transformers torch peft accelerate

Basic Usage

from transformers import T5Tokenizer, T5ForConditionalGeneration
from peft import PeftModel, PeftConfig
import torch

# Load the base model and tokenizer
model_name = "Adilbai/medical-qa-t5-lora"
tokenizer = T5Tokenizer.from_pretrained(model_name)
base_model = T5ForConditionalGeneration.from_pretrained(model_name)

# Load the LoRA configuration and model
config = PeftConfig.from_pretrained(model_name)
model = PeftModel.from_pretrained(base_model, model_name)

def answer_medical_question(question):
    # Prepare the input
    input_text = f"Question: {question}"
    inputs = tokenizer(input_text, return_tensors="pt", max_length=512, truncation=True)
    
    # Generate answer
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_length=256,
            num_beams=4,
            temperature=0.7,
            do_sample=True,
            early_stopping=True
        )
    
    answer = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return answer

# Example usage
question = "What are the symptoms of diabetes?"
answer = answer_medical_question(question)
print(f"Q: {question}")
print(f"A: {answer}")

---

📊 Performance Metrics

🎯 Latest Evaluation Results

Evaluated on: 2025-06-27 15:55:02 by AdilzhanB

Metric	Score	Description
🎯 Exact Match	0.0000	Perfect string matches
📝 Token F1	0.5377	Token-level F1 score
📊 Word Accuracy	0.5455	Word-level accuracy
📏 Length Similarity	0.9167	Response length consistency
🏥 Medical Keywords	0.9167	Medical terminology coverage
⭐ Overall Score	0.5833	Weighted average performance

📈 Performance Highlights

🟢 Excellent Length Similarity (91.67%) - Generates appropriately sized responses
🟢 High Medical Keyword Coverage (91.67%) - Strong medical vocabulary retention
🟡 Good Token F1 Score (53.77%) - Decent semantic understanding
🟡 Moderate Word Accuracy (54.55%) - Room for improvement in precision

---

🔬 Evaluation Results

Test Cases Overview

🧪 Detailed Test Results

Test 1: Perfect Matches ✅

• Samples: 3
• Exact Match: 100%
• Token F1: 100%
• Overall Score: 100%

Test 2: No Matches ❌

• Samples: 3
• Exact Match: 0%
• Token F1: 6.67%
• Overall Score: 20%

Test 3: Partial Matches 🟡

• Samples: 3
• Exact Match: 0%
• Token F1: 66.26%
• Overall Score: 60.32%

Test 4: Medical Keywords 🏥

• Samples: 3
• Medical Keywords: 91.67%
• Overall Score: 58.33%

📝 Sample Comparisons

Example Outputs

Example 1:

• Reference: "Diabetes and hypertension require insulin and medication...."
• Predicted: "Patient has diabetes and hypertension, needs insulin therapy...."
• Token F1: 0.571

Example 2:

• Reference: "Heart disease affects the cardiovascular system significantly...."
• Predicted: "The cardiovascular system shows symptoms of heart disease...."
• Token F1: 0.667

Example 3:

• Reference: "Viral respiratory infections need antiviral treatment, not antibiotics...."
• Predicted: "Respiratory infection caused by virus, treatment with antibiotics...."
• Token F1: 0.375

---

💻 Usage Examples

🔹 Interactive Demo

# Interactive medical Q&A session
def medical_qa_session():
    print("🏥 Medical QA Assistant - Type 'quit' to exit")
    print("-" * 50)
    
    while True:
        question = input("\n🤔 Your medical question: ")
        if question.lower() == 'quit':
            break
            
        answer = answer_medical_question(question)
        print(f"🩺 Answer: {answer}")

# Run the session
medical_qa_session()

🔹 Batch Processing

# Process multiple questions
questions = [
    "What are the side effects of aspirin?",
    "How is pneumonia diagnosed?",
    "What lifestyle changes help with hypertension?"
]

for i, q in enumerate(questions, 1):
    answer = answer_medical_question(q)
    print(f"{i}. Q: {q}")
    print(f"   A: {answer}\n")

---

🛠️ Technical Details

Model Architecture

• Base Model: T5 (Text-to-Text Transfer Transformer)
• Fine-tuning Method: LoRA (Low-Rank Adaptation)
• Parameters: Efficient parameter updates through low-rank matrices
• Training: Supervised fine-tuning on medical QA datasets

Training Configuration

Model: T5 + LoRA
Task: Medical Question Answering
Fine-tuning: Parameter-efficient with LoRA
Evaluation: Multi-metric assessment

---

📚 Citation

If you use this model in your research, please cite:

@model{medical-qa-t5-lora,
  title={Medical QA T5 LoRA: Fine-tuned T5 for Medical Question Answering},
  author={AdilzhanB},
  year={2025},
  url={https://huggingface.co/Adilbai/medical-qa-t5-lora}
}

---

🤝 Contributing

We welcome contributions! Please feel free to:

• 🐛 Report bugs
• 💡 Suggest improvements
• 📊 Share evaluation results
• 🔧 Submit pull requests

---

📄 License

This model is released under the Apache 2.0 License.

---

⚠️ Disclaimer

Important: This model is for educational and research purposes only. It should not be used as a substitute for professional medical advice, diagnosis, or treatment. Always consult with qualified healthcare professionals for medical decisions.

---

Made with ❤️ for the medical AI community

🤗 Hugging Face • 📧 Contact • 🐙 GitHub

Training and evaluation data

keivalya/MedQuad-MedicalQnADataset

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0005
• train_batch_size: 8
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• optimizer: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• num_epochs: 500
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss
2.3794	16.8	50	1.9909
1.2119	33.4	100	0.4473
0.2431	50.0	150	0.0048
0.0343	66.8	200	0.0008
0.0118	83.4	250	0.0003
0.0068	100.0	300	0.0002
0.0042	116.8	350	0.0001
0.0028	133.4	400	0.0001
0.002	150.0	450	0.0000
0.0015	166.8	500	0.0000
0.0012	183.4	550	0.0000
0.0017	200.0	600	0.0000
0.0012	216.8	650	0.0000
0.0008	233.4	700	0.0000
0.0006	250.0	750	0.0000
0.0006	266.8	800	0.0000
0.0004	283.4	850	0.0000
0.0004	300.0	900	0.0000
0.0004	316.8	950	0.0000
0.0004	333.4	1000	0.0000