Ateeqq / news-title-generator

News Title(Headline) Generator 📰

This document details the development of our innovative News Title Generator, designed to produce compelling and informative titles for your news articles. Check the Live Demo Here .

I've tested several other news headline generators on Hugging Face and across the internet, and I can confidently say that this one is the best. 🤗

Model Architecture:

• Foundation: The T5 base model from the Transformers library is our title generator's foundation. This powerful pre-trained model is adept at various text-to-text tasks, making it an ideal choice for our application.
• Fine-Tuning: To optimize performance specifically for news title generation, we fine-tuned the T5 base model on a curated dataset from Hugging Face https://huggingface.co/datasets/Ateeqq/news-title-generator . This dataset consists of over 78,000 training examples, ensuring the model learns the nuances and structure of effective news titles.

How to use?

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

tokenizer = AutoTokenizer.from_pretrained("Ateeqq/news-title-generator")
model = AutoModelForSeq2SeqLM.from_pretrained("Ateeqq/news-title-generator")

def generate_title(input_text):

  input_ids = tokenizer.encode(input_text, return_tensors="pt")
  output = model.generate(input_ids)
  decoded_text = tokenizer.decode(output[0], skip_special_tokens=True)
  return decoded_text

input_text = "A group of scientists discovered a new planet."
generated_title = generate_title(input_text)

print(f"Generated Title: {generated_title}")

Technical Specifications

• Framework: PyTorch, a popular deep learning framework, provides the foundation for our model's development and execution.
• Dataset Split: The training data is strategically divided into two sets: 78,720 examples for training and 19,681 examples for testing. This split ensures the model is adequately trained while reserving a portion for evaluating its generalizability.
• Model Parameters: The fine-tuned model boasts 223 million trainable parameters, allowing it to capture the intricate relationships between text elements that contribute to strong news titles.

Training Configuration

• Batch Size: 8
• Maximum Epochs: The training process iterates through the entire dataset three times (epochs) to ensure thorough learning.
• Global Seed: A fixed random seed (42) is set to guarantee reproducibility of training results.
• Token Length Limits: The source text (article content) is restricted to a maximum of 128 tokens, while the generated titles are capped at 50 tokens.

Key Takeaways

Our News Title Generator leverages the power of the T5 base model, fine-tuned on a comprehensive news title dataset, to deliver exceptional results. The model's architecture and training configuration are meticulously designed to produce high-quality, informative titles within an appropriate character length. This tool empowers creators and journalists to craft impactful headlines that effectively capture readers' attention.

Contact us (at exnrt.com/contact-us ) today to learn more about integrating the News Title Generator into your editorial workflow and unlock the full potential of AI-driven journalism.