Fine-Tune LLAVA Repository

This repository demonstrates the process of fine-tuning LLAVA for various tasks, including data parsing and extracting JSON information from images. It provides comprehensive guidance on how to handle different datasets and fine-tune the model effectively.

---

Video Explanation:

A detailed explanation of the project is available in the following YouTube video:

Fine-Tuning Multimodal LLMs (LLAVA) for Image Data Parsing: Link

---

Repository Structure

Notebooks

• data_exploration/
  Contains notebooks for exploring the Cord-V2 and DocVQA datasets.

• fine-tuning/
  Includes:

  • A notebook for fine-tuning LLAVA 1.6 7B
  • A notebook for testing the fine-tuned model

• test_model/
  Contains multiple notebooks for testing:

  • LLAVA 1.5 7B and 13B
  • LLAVA 1.6 7B, 13B, and 34B

Source Code

• src/
  Contains a Streamlit app to showcase the performance of the fine-tuned model.

  To run the dashboard:

  1. In Terminal 1:

     python src/serve_model.py

  2. In Terminal 2:

     streamlit run src/app.py

  Open the dashboard at http://localhost:8501/ and upload sample images from the data folder to view the results. You can find 20 sample images in the data folder.

---

Installation

1. Clone this repository using:

   git clone https://github.com/Farzad-R/Finetune-LLAVA-NEXT.git

2. Install dependencies from requirements.txt:

   pip install -r requirements.txt

3. Install additional requirements:

   pip install git+https://github.com/huggingface/transformers.git

---

Additional Resources

• Link to Hyperstack Cloud
• HuggingFace Hub to access the model
• A link to a YouTube video will be added here soon to provide further insights and demonstrations.
• LLAVA-NEXT models.
• LLAVA-NEXT info.
• LLAVA-NEXT demo.
• LLAVA-NEXT GitHub repository.
• LLAVA 1.5 demo.
• LLAVA 1.5 GitHub repository.

---