MiniGPT — Lightweight Transformer & Web Application

A lightweight, fully customizable GPT (Decoder-only Transformer) model built from scratch in PyTorch. This project covers the entire end-to-end LLM lifecycle: training a custom BPE tokenizer, training the model, serving inference via a FastAPI backend, and running an interactive Gradio web UI.

---

🏗️ Architecture & Features

• Custom GPT Architecture: Decoder-only Transformer with scaled dot-product Multi-Head Self-Attention, LayerNorm, GELU activations, and positional embeddings.
• Custom Tokenizer: Hugging Face tokenizers Byte-Pair Encoding (BPE) trained specifically on the target dataset.
• REST API: Production-ready FastAPI server (app.py) for serving text generation endpoints.
• Interactive Web UI: Gradio interface (frontend.py) featuring customizable generation parameters (temperature, max tokens).

---

📁 Repository Structure

myGPT/
├── src/
│   ├── model.py        # PyTorch MiniGPT architecture definition
│   ├── train.py        # Tokenizer training & model training loop
│   └── generate.py     # Standalone CLI generation script
├── tokenizer/          # Saved custom tokenizer files (tokenizer.json)
├── app.py              # FastAPI backend API
├── frontend.py         # Gradio web user interface
├── requirements.txt    # Python dependencies
└── README.md