CHFReportGenerator 🫀

An evidence-anchored, research-focused system for automated Congestive Heart Failure (CHF) analysis that provides diagnostic decision support, clinically grounded report generation, and explainable evidence highlighting supporting regions.

🔖 Project at a Glance

• Project name: CHFReportGenerator
• Primary goal: Generate clinically grounded and explainable CHF reports
• Focus: Interpretability, transparency, and reproducibility
• Intended users: Researchers, PhD evaluators, clinicians (research support)

---

✨ Key Features & Contributions

• Evidence-anchored reporting
  Every generated finding is explicitly linked to supporting evidence.
• Clinically grounded narratives
  Outputs are written in structured, clinically meaningful language.
• Parameter-efficient fine-tuning (QLoRA)
  Adapts a large language model with minimal computational cost.
• Research-first design
  Built to support academic evaluation and reproducibility.
• Hardware-efficient
  4-bit quantization enables large-model usage on limited GPU resources.

---

🧠 Model Overview

• Base model: Qwen2.5-VL-7B-Instruct
• Model type: Vision-Language Large Language Model
• Quantization: 4-bit (BitsAndBytes)
• Framework: Unsloth
• Maximum sequence length: 2048 tokens
• Fine-Tuning Method: QLoRA

The base model provides general reasoning and language understanding, while CHF-specific behavior is introduced through lightweight adapters.

⚙️ Installation

Requirements

• Python 3.8 or higher
• CUDA-enabled GPU (recommended)
• PyTorch
• Hugging Face Transformers
• Unsloth
• BitsAndBytes

All dependencies are listed in requirements.txt.

Step-by-Step Setup

git clone https://huggingface.co/aiyubali/CHFReportGenerator
cd CHFReportGenerator
pip install -r requirements.txt