🌟 Overview

A full-stack ChatGPT-like application built (almost) from scratch, featuring real-time conversation capabilities, multi-modal support, and a modern web interface. This project demonstrates the implementation of various components of a production-ready LLM application, from model training to deployment.

✨ Features

• 🤖 Multiple LLM Architecture Support
  • MyGPT (Instruction-tuned GPT2, details below)
  • Gemma-3-1b-it
  • Qwen2.5-0.5B-Instruct
  • SmolVLM-256M-Instruct (Multi-modal)
• 💬 Real-time Conversation
  • WebSocket-based streaming responses
  • Token-by-token generation
• 🎨 Modern Web Interface (vibe coded)
  • React + TypeScript
  • Tailwind CSS for styling
• 🖼️ Multi-modal Capabilities
  • Image upload and processing
  • Vision-language model integration
• 💾 Persistent Storage
  • SQLite database
  • Message and conversation history
• 🐳 Containerization
  • Docker support for both frontend and backend
  • Easy deployment and scaling

🛠️ Technical Stack

Frontend

• React 18
• TypeScript
• Tailwind CSS
• WebSocket for real-time communication

Backend

• FastAPI
• SQLAlchemy with SQLite
• PyTorch
• Transformers
• WebSockets
• Docker

🚀 Getting Started

Installation

1. Clone the repository

git clone https://github.com/snnclsr/chatgpt-from-scratch.git
cd chatgpt-from-scratch

2. With docker-compose

cd chatgpt-from-scratch
docker-compose up --build

3. or separately without the docker

Backend:

uvicorn backend.main:app --host 0.0.0.0 --port 8000 --reload

Frontend:

cd frontend
npm start

📚 Model Training

The training procedure is under modelling directory building upon Sebastian Raschka's "Build a Large Language Model (From Scratch)" book, adapted from here and implementing:

• GPT2 model architecture
• Instruction tuning using the Alpaca dataset (from the pretrained weights)

To run the training:

python -m modelling.train

I also applied following changes to the training code/model to make training/inference faster (https://github.com/rasbt/LLMs-from-scratch/tree/main/ch05/10_llm-training-speed)

1. Create causal mask on the fly
2. Use tensor cores
3. Fused AdamW optimizer
4. Replacing from-scratch code by PyTorch classes
5. Using FlashAttention
6. Using pytorch.compile
7. Increasing the batch size

As outlined by Sebastian as well, these updates make everything go faster, 6-7 times.

🙏 Acknowledgments

• Sebastian Raschka's "Build a Large Language Model (From Scratch)" book
• Alpaca dataset (CC BY-NC 4.0)
• Open-source model providers (Gemma-3-1b-it, Qwen2.5-0.5B-Instruct, SmolVLM-256M-Instruct)

Future Enhancements

• Markdown support