AventIQ-AI/gpt2-news-article-generation

News Article Generation with GPT2

This repository hosts a quantized version of the GPT2 model, fine-tuned for generation of news article tasks. The model has been optimized for efficient deployment while maintaining high accuracy, making it suitable for resource-constrained environments.

Model Details

• Model Architecture: gpt2
• Task: Text Summarization
• Dataset: Hugging Face's `ag_news'
• Quantization: Float16
• Fine-tuning Framework: Hugging Face Transformers

Usage

Installation

pip install transformers torch

Loading the Model

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"

model_name = "AventIQ-AI/gpt2-news-article-generation"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name).to(device)

import torch
import html

# Define test text
test_text = "The future of AI"

# Tokenize input
inputs = tokenizer(test_text, return_tensors="pt").to("cuda")

# Generate response
with torch.no_grad():
    output_tokens = model.generate(
        **inputs,
        max_length=200,  # Allow longer responses
        num_beams=5,  # Balances quality & diversity
        repetition_penalty=2.0,  # Reduce repeating patterns
        temperature=0.2,  # More deterministic response
        top_k=100,  # Allows more diverse words
        top_p=0.9,  # Keeps probability confidence
        do_sample=True,  # Sampling for variety
        no_repeat_ngram_size=3,  # Prevents excessive repetition
        num_return_sequences=1,  # Returns one best sequence
        early_stopping=True,  # Stops when response is complete
        length_penalty=1.2,  # Balances response length
        pad_token_id=tokenizer.eos_token_id,  # Prevents truncation
        eos_token_id=tokenizer.eos_token_id,  # Ensures completion
        return_dict_in_generate=True,  # Structured output
        output_scores=True  # Debugging purposes
    )

# Decode and clean response
generated_response = tokenizer.decode(output_tokens.sequences[0], skip_special_tokens=True)
cleaned_response = html.unescape(generated_response).replace("#39;", "'").replace("quot;", '"')

print("\nGenerated Response:\n", cleaned_response)

📊 ROUGE Evaluation Results

After fine-tuning the T5-Small model for text summarization, we obtained the following ROUGE scores:

Metric	Score	Meaning
ROUGE-1	0.3061 (~30%)	Measures overlap of unigrams (single words) between the reference and generated summary.
ROUGE-2	0.1241 (~12%)	Measures overlap of bigrams (two-word phrases), indicating coherence and fluency.
ROUGE-L	0.2233 (~22%)	Measures longest matching word sequences, testing sentence structure preservation.
ROUGE-Lsum	0.2620 (~26%)	Similar to ROUGE-L but optimized for summarization tasks.

Fine-Tuning Details

Dataset

The Hugging Face's ag_news dataset was used, containing the text and their labels.

Training

• Number of epochs: 3
• Batch size: 4
• Evaluation strategy: epoch
• Learning rate: 5e-5

Quantization

Post-training quantization was applied using PyTorch's built-in quantization framework to reduce the model size and improve inference efficiency.

Repository Structure

.
├── model/               # Contains the quantized model files
├── tokenizer_config/    # Tokenizer configuration and vocabulary files
├── model.safetensors/   # Quantized Model
├── README.md            # Model documentation

Limitations

• The model may not generalize well to domains outside the fine-tuning dataset.
• Quantization may result in minor accuracy degradation compared to full-precision models.

Contributing

Contributions are welcome! Feel free to open an issue or submit a pull request if you have suggestions or improvements.