# Text-to-Text Transfer Transformer Quantized Model for News Summarization

This repository hosts a quantized version of the T5 model, fine-tuned specifically for text summarization of news. The model extracts concise summaries from semi-structured or unstructured news texts, making it ideal for POS systems, kitchen displays, and chat-based food order logging.

## Model Details

  - **Field:**	Description
  - **Model Architecture** T5 (Text-to-Text Transfer Transformer)
  - **Task**	Text Summarization for News
  - **Input Format**	Free-form order text (includes Order ID, Customer, Items, etc.)
  - **Quantization**	8-bit (int8) using bitsandbytes
  - **Framework**	Hugging Face Transformers
  - **Base Model**	t5-base
  - **Dataset**   Custom

## Usage

## Installation

```sh
pip install transformers accelerate bitsandbytes torch
```
### Loading the Model

```python
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
import torch

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

model_name = "AventIQ-AI/T5-News-Summarization"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name, load_in_8bit=True, device_map="auto")

def test_summarization(model, tokenizer):
    user_text = input("\nEnter your News text:\n")
    inputs = tokenizer("summarize: " + user_text, return_tensors="pt", truncation=True, max_length=512).to(model.device)

    output = model.generate(
        **inputs,
        max_new_tokens=100,
        num_beams=5,
        length_penalty=0.8,
        early_stopping=True
    )

    summary = tokenizer.decode(output[0], skip_special_tokens=True)
    return summary

print("\n📝 **Model Summary:**")
print(test_summarization(model, tokenizer))
```

## ROUGE Evaluation Results

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

| **Metric**  | **Score**  | **Meaning** |
|-------------|-----------|-------------|
| **ROUGE-1**  | **0.4125** (~41%)	| Overlap of **unigrams** between reference and summary. |
| **ROUGE-2**  | **0.2167** (~22%)	| Overlap of **bigrams**, indicating fluency. |
| **ROUGE-L**  | **0.3421** (~34%)	| Longest common subsequence matching structure. |
| **ROUGE-Lsum**  | **0.3644** (~36%) | Sentence-level summarization effectiveness. |

## Fine-Tuning Details

### Dataset

Custom-labeled food order dataset containing fields like Order ID, Customer, and Order Details. The model was trained to extract clean, natural summaries from noisy or inconsistent order formats.

### Training

  - Number of epochs: 3

  - Batch size: 4

  - Evaluation strategy: epoch

  - Learning rate: 3e-5

### Quantization

Post-training 8-bit quantization using bitsandbytes library with Hugging Face integration. This reduced the model size and improved inference speed with negligible impact on summarization quality.

## Repository Structure

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


## Limitations

 - The model may misinterpret or misformat input with excessive noise or missing key fields.

 - Quantized versions may show slight accuracy loss compared to full-precision models.

 - Best suited for English-language food order formats.  

## Contributing

Contributions are welcome! If you have suggestions, feature requests, or improvements, feel free to open an issue or submit a pull request.