hf-summarization-pipeline

Summarization Pipeline Project

This project demonstrates how to use the Hugging Face Pipeline API to load a pre-trained summarization model and generate summaries from example text.

What Was Done

1. Model Selection: Used facebook/bart-large-cnn, a pre-trained BART (Bidirectional and Auto-Regressive Transformers) model that was fine-tuned on the CNN/DailyMail dataset for abstractive summarization.

2. Pipeline API: Leveraged Hugging Face's pipeline API, which provides a simple interface for:

   • Loading pre-trained models
   • Handling tokenization automatically
   • Running inference with minimal code

3. Examples: Tested the model on multiple example texts covering different topics:

   • Artificial intelligence and machine learning
   • Climate change
   • The Renaissance period
   • A longer article-style text

4. Configuration: Used parameters to control summary length:

   • max_length: Maximum number of tokens in the summary
   • min_length: Minimum number of tokens in the summary
   • do_sample: Set to False for deterministic (greedy) decoding

Files

• summarization_pipeline.py: Main script demonstrating the pipeline API
• README.md: This file

Usage

Installation

pip install transformers torch

Running the Scripts

Option 1: Simple version (recommended, avoids dependency issues)

python simple_summarization.py

Option 2: Pipeline API version

python summarization_pipeline.py

Note: May have dependency issues in some environments. See SETUP_GUIDE.md

Option 3: Using Inference API

pip install requests
python summarization_api.py

Using in Your Own Code

from transformers import pipeline

# Load the summarization pipeline
summarizer = pipeline("summarization", model="facebook/bart-large-cnn")

# Summarize text
text = "Your long article or text here..."
summary = summarizer(text, max_length=100, min_length=30)
print(summary[0]['summary_text'])

Model Information

• Model: facebook/bart-large-cnn
• Task: Abstractive Summarization
• Training Data: CNN/DailyMail dataset
• Architecture: BART (Bidirectional and Auto-Regressive Transformers)

Alternative Models

You can easily switch to other pre-trained summarization models:

• google/pegasus-xsum: Trained on XSum dataset
• t5-small: General-purpose T5 model
• sshleifer/distilbart-cnn-12-6: Smaller, faster version

Simply change the model parameter in the pipeline:

summarizer = pipeline("summarization", model="google/pegasus-xsum")

Notes

• The model requires significant memory (the large BART model is ~1.6GB)
• For GPU acceleration, change device=-1 to device=0 in the pipeline
• The first run will download the model, which may take some time