T5 Finetuning for SQuAD Question Answering

This project demonstrates how to finetune a T5 model on the Stanford Question Answering Dataset (SQuAD) for question answering tasks. The code includes data loading, model training, and an interactive QA system.

Python 3.7+

pip (Python package installer)

Clone this repository:

Install the required packages:

This project uses the Stanford Question Answering Dataset (SQuAD) v1.1. You need to download the dataset files and place them in the project root directory:

Download train-v1.1.json

Download dev-v1.1.json

main.py: The main script to run the entire pipeline

data_loading.py: Functions for loading and preprocessing the SQuAD dataset

model_utils.py: Utility functions for initializing the model and tokenizer

training.py: Functions for training the model

qa_utils.py: Utilities for question answering, including an interactive QA system

To run the entire pipeline (data loading, model training, and interactive QA):

This will:

Load the SQuAD dataset

Initialize the T5 model and tokenizer

Train the model on the SQuAD dataset

Save the trained model

Test the model with a sample question

Start an interactive QA session

Data Loading: The load_squad function in data_loading.py reads the SQuAD JSON files and converts them into a format suitable for training.

Model Initialization: The initialize_model_and_tokenizer function in model_utils.py loads the pretrained T5 model and tokenizer.

Training: The train_model function in training.py sets up the training arguments, tokenizes the dataset, and trains the model using the Hugging Face Trainer class.

Question Answering: The answer_question function in qa_utils.py takes a question and context, processes them through the model, and returns the generated answer.

Interactive QA: The interactive_qa function in qa_utils.py provides a command-line interface for users to input contexts and questions, and receive answers from the model.

Contributions are welcome! Please feel free to submit a Pull Request.

This project is open source and available under the MIT License.