Local MCQ Generation Chatbot

This project is a streamlit-based chatbot interface built on top of a fine-tuned LLaMA 3.2B model, designed to generate high-quality multiple-choice questions (MCQs) from user prompts and a given PDF documents. The model is fine-tuned specifically to target Computer Science education use cases and provides responses in a well-formatted, exam-style MCQ format.

---

Features

• Locally hosted LLM (LLaMA 3.2B fine-tuned) for fast, private inference.
• ChatGPT-like frontend using Streamlit, with left-right conversational UI.
• PDF upload support: You can attach a document and generate MCQs based on its contents.
• Streaming response: Output is streamed character-by-character for better user experience.
• Lazy model loading: Shows a loading screen while the model initializes, avoiding UI freeze.
• Custom MCQ generation format based on user prompt and document context.

---

Usage

1. Clone the Repository

git clone https://github.com/your-username/local-mcq-chatbot.git
cd local-mcq-chatbot

2. Setup Virtual Environment

python -m venv mcq_gen
source mcq_gen/bin/activate     # On Windows: mcq_gen\Scripts\activate

3. Install Requirements

pip install -r requirements.txt

4. Create .env File

Inside your root directory, create a .env file and add:

hf_token=your_huggingface_access_token

5. Run the App

streamlit run main.py

---

Model Info

• Model: sinister007/llama-3.2-1b-mcq-gen

• Fine-tuned on: Hand-curated MCQ-style datasets focused on Computer Science.

• Format follows:

  Question: ...
  A. ...
  B. ...
  C. ...
  D. ...
  Answer: X
  

---

Example Use

• Ask: Generate MCQs on Operating System deadlock concepts.
• Upload: A PDF textbook or lecture note.
• Get: A list of 3–5 well-formed, challenging MCQs generated in real time.

---

Project Structure

.
├── frontend.py              # Streamlit UI
├── inference.py             # Model loading and generation
├── text_extracter.py        # PDF text extraction logic
├── requirements.txt
└── .env                     # Environment variables (not committed)

---

Roadmap

• Further fine-tune the model on a richer, domain-specific dataset.
• Integrate retrieval-based augmentation (RAG) for better factual grounding.
• Add support for answer explanations and difficulty tagging.
• Export MCQs as JSON or PDF.