🌟 Summarization Streamlit Application Link

📝 KDTS Task 04 — Text Summarization Using NLP

Welcome to the repository for Task KDT04: Text Summarization Using NLP, where we fine-tune a transformer-based model to generate high-quality abstractive summaries using Hugging Face Transformers and the CNN/DailyMail dataset.

---

🎯 Objective

Train or fine-tune a transformer model for text summarization using T5, BART, or PEGASUS, and demonstrate the ability to generate concise summaries for long-form articles or news content.

---

✅ Requirements Checklist

Requirement	Status
Use T5, BART, or PEGASUS from Hugging Face	✅ T5-small
Train on a small custom dataset or CNN/DailyMail	✅ 60K rows from CNN/DailyMail
Generate summaries for at least 5 sample paragraphs	✅ 10+ samples
Compare model-generated summaries vs. extractive method	✅ Compared generated vs extractive summary
Include a web form to paste text and view summary	✅ Created and deployed a streamlit application on hugging face

---

📚 Dataset

We used the CNN/DailyMail dataset and combined 60,000 samples from 3 versions:

• 📘 20K from version 1.0.0
• 📙 20K from version 2.0.0
• 📗 20K from version 3.0.0

Split Configuration:

• 🔹 Train: 38,400 samples
• 🔹 Validation: 9,600 samples
• 🔹 Test: 12,000 samples

---

🤖 Model Used

We fine-tuned the t5-small model for summarization and published it on the Hugging Face Model Hub:

🔗 Model on Hub: trohith89/KDTS_T5_Summary_FineTune

---

🛠️ Project Pipeline

1. Data Preparation

• Loaded 3 versions of CNN/DailyMail using Hugging Face Datasets
• Combined them into a single DatasetDict
• Split into train/validation/test

2. Tokenization

• Prefixed each article with summarize:
• Used T5TokenizerFast with truncation and padding

3. Model Fine-Tuning

• Used Seq2SeqTrainer from Hugging Face with:
  • Learning rate: 3e-4
  • Epochs: 3
  • Batch size: 4
  • Metric: rougeL for best model checkpoint
  • Early stopping and mixed-precision (FP16) enabled

4. Inference

• Generated summaries for test samples
• Compared model summaries with reference summaries

---

💡 How to Use the Model

✅ Option 1: Hugging Face pipeline (Easy Inference)

from transformers import pipeline

pipe = pipeline("text2text-generation", model="trohith89/KDTS_T5_Summary_FineTune")

text = "summarize: The Indian economy grew 6.1% last quarter due to..."
summary = pipe(text)[0]['generated_text']
print("Summary:", summary)

✅ Option 2: Load Tokenizer and Model Manually

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("trohith89/KDTS_T5_Summary_FineTune") model = AutoModelForSeq2SeqLM.from_pretrained("trohith89/KDTS_T5_Summary_FineTune")

input_text = "summarize: India’s economy grew 6.1% last quarter..." inputs = tokenizer(input_text, return_tensors="pt", truncation=True) outputs = model.generate(**inputs) summary = tokenizer.decode(outputs[0], skip_special_tokens=True) print("Summary:", summary)

📈 Evaluation Metrics We used the following metrics for evaluation:

ROUGE-1

ROUGE-2

ROUGE-L

ROUGE-Lsum

These metrics compare the overlap between the generated and reference summaries.

KDTS_Task4_Summarization/ ├── KDTS_TASK4_FineTuning_Summary_TR.ipynb # Full training and inference pipeline ├── KDTS-outputs.txt # Code outputs from sample runs ├── requirements.txt # Required dependencies ├── README.md # This documentation ├── app.py # ⏳ Streamlit app (Coming soon)

🙌 Acknowledgements 🤗 Hugging Face Transformers

📰 CNN/DailyMail Dataset

🧪 KDTS Internship Program Task Series

📬 Contact Have questions or feedback? 📧 Reach out via [trohith89@gmail.com]

⭐️ Star this repo if you found it useful!