WhoRitikrana/ContextIQ

👨‍💻 Author

Ritik Rana
NLP Engineer | AI Developer | RAG Engineer contact : ritikrana056@gmail.com

📘 Question Answering System using BERT & SQuAD

An end-to-end Question Answering (QA) system built using Hugging Face Transformers, trained on the SQuAD dataset, and deployed with a Streamlit web app.

This project demonstrates:

• Fine-tuning BERT for Extractive Question Answering
• Custom preprocessing with offset mapping
• Hugging Face Trainer API (v5 compatible)
• Model saving & loading
• Streamlit deployment

---

🚀 Project Overview

This project trains a BERT-based model to answer questions from a given paragraph (context).

Example:

Context:

TensorFlow is an open-source machine learning framework developed by Google.

Question:

Who developed TensorFlow?

Answer:

Google

The model predicts:

• Start token index
• End token index

And extracts the correct span from the context.

---

🧠 Model Architecture

• Base Model: bert-base-uncased
• Task: Extractive Question Answering
• Dataset: SQuAD v1
• Framework: Hugging Face Transformers
• Training API: Trainer
• Deployment: Streamlit

---

📂 Project Structure

qa_squad_project/
│
├── config.py               # Configuration file
├── data_loader.py          # Loads SQuAD dataset
├── preprocess.py           # Tokenization + offset mapping
├── model.py                # Loads QA model
├── train.py                # Training script
├── inference.py            # Inference script
├── streamlit_app.py        # Web application
├── qa_model/               # Saved trained model
└── README.md

Training dataset

Squad dataset

Model initial version train on only

2000 training dataset and 500 validation dataset on 3 epochs 

🏋️ Training the Model

Run:

python train.py

Training will:

• Load SQuAD dataset
• Tokenize using offset mapping
• Fine-tune BERT
• Save model in qa_model/

---

🧪 Run Inference (CLI)

After training:

python inference.py

This will test the model with a sample question.

---

🌐 Run Streamlit Web App

streamlit run streamlit_app.py

If using remote/cloud environment:

streamlit run streamlit_app.py --server.address 0.0.0.0 --server.port 8501

Open browser at:

http://localhost:8501

---

🔍 How It Works

1️⃣ Tokenization

• Question and context are tokenized together.
• Offset mapping maps tokens back to original character positions.

2️⃣ Label Creation

• Convert answer character positions → token indices.
• Save as:
  • start_positions
  • end_positions

3️⃣ Training

Model learns to predict:

• Start logits
• End logits

Loss is computed using CrossEntropyLoss.

4️⃣ Inference

Model predicts best start & end token span. Pipeline extracts text from context.

---

🧪 Example Output

Question: Who developed TensorFlow?
Answer: Google
Confidence: 0.98

---

💾 Saved Model Files

Inside qa_model/:

File	Purpose
pytorch_model.bin	Trained weights
config.json	Model architecture
tokenizer.json	Tokenizer rules
training_args.bin	Training config

Load later using:

from transformers import AutoModelForQuestionAnswering, AutoTokenizer

model = AutoModelForQuestionAnswering.from_pretrained("qa_model")
tokenizer = AutoTokenizer.from_pretrained("qa_model")

---

📈 Future Improvements

• Train on full SQuAD dataset
• Add SQuAD v2 (no-answer support)
• Add answer highlighting in Streamlit
• Add PDF upload support
• Convert to RAG system
• Deploy on Hugging Face Spaces

---

🖥 Hardware Used

Recommended GPU:

• NVIDIA T4 (16GB)

Works on CPU for small datasets (slower).

---

---

license apache-2.0