📝 SummarizeIt — AI-Powered Summarization App

A full-stack AI app that summarizes long conversations and articles using a fine-tuned Pegasus transformer. Includes chat vs. paragraph detection, summary length control, REST API, and clean responsive UI.

Tech Stack: Python • FastAPI • Transformers • PyTorch • Bootstrap • Docker

🎯 Try it Live • 🤖 Model • 🐛 Report Bug

📖 Inspired by recent research on fine-tuning Pegasus for dialogue summarization
(see this paper)

---

✨ Features

• 🧠 Fine-tuned Pegasus Model - Trained on SAMSum for summarizing dialogue and paragraphs
• 🤗 Hugging Face Integration - Model deployed on Hugging Face Hub, app hosted on Spaces
• 🎯 Smart Input Detection - Automatically detects chat vs. paragraph format
• 📏 Multiple Summary Lengths - Short, medium, and long summary options
• 🌐 Modern Web Interface - Clean, responsive design with real-time processing
• 📁 File Upload Support - Handle .txt and .md files up to 10MB
• 🔌 RESTful API - Complete API with OpenAPI documentation
• 📊 ROUGE Evaluation - Comprehensive model performance metrics

---

🔄 How It Works

📊 Complete Workflow

graph TD
    A[📚 SAMSum Dataset] --> B[🔧 Data Preprocessing]
    B --> C[🧠 Fine-tune Pegasus Model]
    C --> D[📈 Model Evaluation]
    D --> E[💾 Save Fine-tuned Model]
    E --> F[🤗 Upload to Hugging Face Hub]
    F --> G[🚀 Deploy FastAPI App]
    G --> H[🌐 Hugging Face Spaces]
    
    I[👤 User Input] --> J{📝 Input Type Detection}
    J -->|Chat| K[💬 Chat Processing]
    J -->|Article| L[📄 Article Processing]
    K --> M[🤖 Pegasus Model]
    L --> M
    M --> N[📋 Generate Summary]
    N --> O[📊 Calculate Metrics]
    O --> P[✅ Return Results]

Loading

🎯 Processing Pipeline

1. 📥 Input Processing

   • Detect input type (chat conversation vs. article)
   • Clean and preprocess text
   • Handle file uploads (.txt, .md)

2. 🧠 AI Summarization

   • Load fine-tuned Pegasus model from Hugging Face Hub
   • Tokenize input text (max 1024 tokens)
   • Generate summary based on selected length
   • Apply post-processing filters

3. 📊 Output Generation

   • Calculate compression ratio
   • Compute summary statistics
   • Format response with metadata
   • Return JSON response or web interface

🔧 Model Training Pipeline

graph LR
    A[📚 SAMSum Dataset<br/>16k+ conversations] --> B[🔧 Tokenization<br/>Max 1024 tokens]
    B --> C[🎯 Fine-tuning<br/>4 epochs]
    C --> D[📈 ROUGE Evaluation<br/>R-1, R-2, R-L]
    D --> E[💾 Model Export<br/>HuggingFace format]

Loading

Training Stats:

• Dataset: 16,000+ chat conversations
• Training Time: 2-4 hours (GPU) / 8-12 hours (CPU)
• Model Size: ~2.3GB
• Performance: 11.9% improvement in ROUGE-1 score

---

🚀 Quick Start

🎯 Try Online (No Installation Required)

💻 Local Setup

Option 1: Manual Setup

git clone https://github.com/ananthakr1shnan/SummarizeIt.git
cd SummarizeIt
python -m venv venv
source venv/bin/activate  # or venv\Scripts\activate on Windows
pip install -r requirements.txt
python main.py setup
python main.py serve

Option 2: Docker

git clone https://github.com/ananthakr1shnan/SummarizeIt.git
cd SummarizeIt
docker-compose up --build

🌐 Access the app at: http://localhost:8000

---

📋 Requirements

Tech Stack:

• Backend: Python 3.11+, FastAPI, Uvicorn
• AI/ML: Transformers, PyTorch, Datasets
• Frontend: Bootstrap, HTML/CSS/JavaScript
• Deployment: Docker, Hugging Face Spaces
• Model: Fine-tuned Pegasus (google/pegasus-cnn_dailymail)

System Requirements:

• 8GB+ RAM (for model training)
• CUDA GPU (optional, for faster training)
• Docker (optional)

---

🎯 Usage

Web Interface

1. Open http://localhost:8000
2. Paste text or upload a file
3. Select summary length and content type
4. Click "Generate Summary"
5. Copy and share your summary

API Usage

import requests

response = requests.post("http://localhost:8000/summarize", json={
    "text": "Your long text here...",
    "summary_length": "medium",
    "input_type": "auto"
})

result = response.json()
print(result["summary"])

CLI Commands

python main.py setup      # Setup environment
python main.py train      # Train model (2-4 hours)
python main.py serve      # Start web server
python main.py evaluate   # Evaluate model

---

🧠 Model Performance

ROUGE Scores (Fine-tuned vs Base)

Metric	Base Model	Fine-tuned	Improvement
ROUGE-1	0.42	0.47	+11.9%
ROUGE-2	0.19	0.23	+21.1%
ROUGE-L	0.34	0.39	+14.7%

⚠️ Note: Due to hardware constraints, the model was trained for 1 epoch.
Full training with 4 epochs and a high-performance GPU would significantly improve results.

Performance Stats

• Processing Speed: ~2-3 seconds per summary
• Memory Usage: ~2GB GPU / ~4GB CPU
• Batch Processing: ~1 second per text in batch

🔌 API Endpoints

Method	Endpoint	Description
POST	/summarize	Summarize single text
POST	/summarize/batch	Summarize multiple texts
POST	/summarize/file	Upload and summarize file
GET	/health	API health check
GET	/docs	Interactive API documentation

Request Format

{
  "text": "Your text here...",
  "summary_length": "medium",  // "short", "medium", "long"
  "input_type": "auto"         // "auto", "chat", "paragraph"
}

Response Format

{
  "summary": "Generated summary text...",
  "original_length": 450,
  "summary_length": 89,
  "compression_ratio": 0.2,
  "summary_type": "medium",
  "input_type": "paragraph"
}

---

🎓 Training

Train your own model with custom data:

# Quick training with default settings
python main.py train

# Custom configuration in config/settings.py
NUM_EPOCHS = 4 # Use 4 for full training. Set to 1 if you're working with limited hardware
BATCH_SIZE = 2
LEARNING_RATE = 5e-5

The training process includes:

1. Download SAMSum dataset
2. Preprocessing and tokenization
3. Model training
4. ROUGE evaluation
5. Model saving

---

🚀 Deployment

🤗 Hugging Face Deployment

• Live App: Deployed on Hugging Face Spaces
• Model Hub: Fine-tuned model available at Hugging Face Hub
• Zero Setup: No installation required, just click and use!

---

📚 Reference

This project draws inspiration from:

Fine-tuning the Large Language Pegasus Model for Dialogue Summarization

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

🙋‍♂️ About Me

Hi, I'm Ananthakrishnan K — a B.Tech CS student passionate about Machine Learning and AI.
This project is part of my journey to master NLP and build real-world AI tools.

📫 Email • LinkedIn •