---
base_model: microsoft/Florence-2-base
library_name: peft
tags:
  - medical
  - vision-language-model
  - vqa
  - radiology
---

# Generative AI Radiology VLM (Florence-2)

This model is a Parameter-Efficient Fine-Tuned (PEFT/LoRA) version of Microsoft's `Florence-2-base`. It has been specifically trained on the **VQA-RAD** dataset to act as a Generative AI Vision-Language Model capable of answering free-form textual questions about medical X-Rays.

## Model Details

- **Architecture**: Vision Encoder + Text Decoder (Florence-2)
- **Task**: Medical Visual Question Answering (VQA)
- **Fine-Tuning Technique**: Low-Rank Adaptation (LoRA)
- **Target Modules**: `q_proj`, `v_proj`, `o_proj`

## Training Results

The model was fine-tuned for 3 epochs on an NVIDIA A100-40GB GPU using mixed precision (fp16). The training loss steadily decreased, demonstrating strong anatomical and vocabulary convergence.

![Training Loss](vlm_loss_curve.png)

## Local Web UI (Gradio)

The repository includes a local `app.py` script that loads these LoRA adapters and spins up a local web UI for inference.

![Gradio Web UI Demo](ScreenRecord.gif)

### Framework versions

- PEFT 0.11.1
- Transformers 4.42.4