LED Document Summarization App

A Streamlit application that summarizes documents using the Longformer Encoder-Decoder (LED) model, specifically the "pszemraj/led-large-book-summary" model which is fine-tuned for book summarization.

Features

• PDF document upload and display
• Leverages the LED model to handle long documents (up to 16,384 tokens)
• Optimized for book and long-form content summarization
• Configurable summarization parameters
• Progress indicators during processing
• Error handling and logging
• Summary download option
• Responsive UI with sidebar configuration

Requirements

• Python 3.11 or later
• PyTorch
• Transformers
• LangChain
• Streamlit
• See requirements.txt for complete dependencies

Installation

1. Clone this repository:

git clone https://github.com/yourusername/document-summarization-app.git
cd document-summarization-app

2. Create and activate a virtual environment:

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install the required packages:

pip install -r requirements.txt

4. Download the LED model:

mkdir -p models
python download_model.py

# Or use the model directly from HuggingFace by setting
# MODEL_PATH=pszemraj/led-large-book-summary in your .env file

5. Configure the application:

cp .env.example .env
# Edit .env file to match your environment

Usage

1. Run the Streamlit app:

streamlit run app.py

2. Open your browser at http://localhost:8501

3. Upload a PDF document and configure summarization parameters in the sidebar

4. Click "Summarize Document" to generate a summary

Model Configuration

The application is configured to use the LED Large Book Summary model by default. This model:

• Handles up to 16,384 tokens of input
• Is fine-tuned on the BookSum dataset
• Is optimized for long-form content like books, articles, and reports

To add or change models:

1. Add your model to the models directory or specify its path in the .env file
2. Update the AVAILABLE_MODELS dictionary in tools.py with your new model

Optimization Notes

For best results with the LED model, the application uses these parameters:

• no_repeat_ngram_size=3
• encoder_no_repeat_ngram_size=3
• repetition_penalty=3.5
• num_beams=4
• early_stopping=True

These settings help create more coherent and less repetitive summaries.

Project Structure

• app.py: Main Streamlit application
• tools.py: Model loading and text processing utilities
• requirements.txt: Package dependencies
• Dockerfile: Container configuration
• .env.example: Example environment configuration
• data/: Directory for uploaded documents (created automatically)
• models/: Directory for model files

License

MIT License