iiCEMAN/BanglaMedVQA

Medical VQA LoRA Training

Fine-tuning google/medgemma-4b-it vision-language model using FP32 LoRA with LLaMA Factory for Bengali medical question answering.

📁 Repository Structure

med-vqa-lora/
├── configs/
│   ├── qwen2.5-vl-7b.yaml        # Training configuration
│   ├── qwen2.5-vl-7b-test.yaml   # Testing configuration
│   ├── med-gemma-4b.yaml         # Legacy MedGemma training config
│   └── med-gemma-4b-test.yaml    # Legacy MedGemma testing config
├── data/
│   ├── dataset_info.json          # Master dataset registry
│   ├── chest_x-ray/
│   │   ├── train/
│   │   │   ├── images/            # Training images
│   │   │   ├── chest_x-ray.csv    # Training data
│   │   │   └── chest_x-ray_dataset.json
│   │   └── test/
│   │       ├── images/            # Test images
│   │       ├── chest_x-ray.csv    # Test data
│   │       └── chest_x-ray_dataset.json
│   └── medicat/
│       └── train/
│           ├── images/            # Medicat training images
│           ├── medicat.csv        # Medicat training data
│           └── medicat_dataset.json
├── output/                        # Model outputs and predictions
├── prepare_data.py                # Data preparation script
├── train_model.py                 # Main training/testing script
├── requirements.txt               # Python dependencies
├── Dockerfile                     # Container setup
└── README.md                      # This file

🚀 Usage

1. Build Docker Container

docker build -t med-vqa:latest .

2. Prepare Data

# Prepare all datasets
docker run --rm -v $(pwd):/app med-vqa:latest python prepare_data.py --all

# Prepare specific dataset
docker run --rm -v $(pwd):/app med-vqa:latest python prepare_data.py --dataset chest_x-ray --split train

3. Run Training

docker run --gpus all --ipc=host --rm --env-file .env -v $(pwd):/app med-vqa:latest python train_model.py --config configs/med-gemma-4b.yaml

4. Run Testing

docker run --gpus all --ipc=host --rm --env-file .env -v $(pwd):/app med-vqa:latest python train_model.py --config configs/med-gemma-4b-test.yaml

5. Complete Pipeline

# Train and test in sequence
docker run --gpus all --ipc=host --rm --env-file .env -v $(pwd):/app med-vqa:latest python train_model.py --config configs/med-gemma-4b.yaml
docker run --gpus all --ipc=host --rm --env-file .env -v $(pwd):/app med-vqa:latest python train_model.py --config configs/med-gemma-4b-test.yaml

⚙️ Configuration

Training (configs/med-gemma-4b.yaml)

• Model: google/medgemma-4b-it
• LoRA Rank: 16, Alpha: 32
• Batch Size: 2 per device, 2 gradient accumulation
• Template: gemma3

Testing (configs/med-gemma-4b-test.yaml)

• Loads trained adapter from ./output
• Generates predictions for test dataset

📝 Data Preparation

# Prepare all datasets
python prepare_data.py --all

# Prepare specific dataset
python prepare_data.py --dataset chest_x-ray --split train

# Custom paths
python prepare_data.py --dataset chest_x-ray --split test --csv data/chest_x-ray/test/chest_x-ray.csv --images data/chest_x-ray/test/images --output data/chest_x-ray/test/chest_x-ray_dataset.json

🎯 Output Files

Training Outputs

• output/adapter_model.safetensors - LoRA weights (119MB)
• output/adapter_config.json - LoRA configuration
• output/checkpoint-32/ - Training checkpoint
• output/train_results.json - Training metrics

Test Predictions

• output/chest_x-ray_test_predictions.csv - Test predictions with columns:
  • image_id, image_path, category, category_bn
  • question, question_bn, llm_answer, llm_answer_bn
  • predicted_answer_bn - Model's predictions

🔍 Example Usage

Sample Predictions

image_id,image_path,question_bn,llm_answer_bn,predicted_answer_bn
test_001,data/chest_x-ray/test/images/00015953_015.png,এখানে কোন নির্দিষ্ট অবস্থা চিহ্নিত করা হয়েছে?,কোন নির্দিষ্ট অবস্থা চিহ্নিত হয়নি।,কোন নির্দিষ্ট অবস্থা চিহ্নিত করা হয়নি।
test_002,data/chest_x-ray/test/images/00011237_094.png,অনুপ্রবেশটি কোথায় অবস্থিত?,মধ্য ডান,মধ্য বাম

Model Loading

from transformers import AutoModelForCausalLM, AutoProcessor
from peft import PeftModel

model = AutoModelForCausalLM.from_pretrained("google/medgemma-4b-it")
processor = AutoProcessor.from_pretrained("google/medgemma-4b-it")
model = PeftModel.from_pretrained(model, "./output")