README.md

---
title: Formula Engine Chatbot
emoji: 🧮
colorFrom: purple
colorTo: blue
sdk: gradio
sdk_version: 5.32.0
app_file: app.py
pinned: false
license: apache-2.0
tags:
- ml-intern
---

# 🧮 Formula Engine Chatbot

## AI-Powered Weight Compression: Qwen 0.5B from Mathematical Formulas

This Space demonstrates the **Formula Engine** concept:
- Instead of storing the full **942 MB** Qwen 0.5B model
- We store **compact mathematical formulas** (~474 MB) that reconstruct the weights on-the-fly
- **49.7% disk space saved** while maintaining 99.99% accuracy

### How it works:
1. **Formula Discovery**: Analyzes each weight matrix to find the most compact representation
2. **4-bit Quantization**: `W ≈ scale × W_q + zero_point` (75% compression per layer)
3. **SVD Factorization**: `W ≈ U_r × S_r × V_r^T` (for rectangular matrices)
4. **On-the-fly Reconstruction**: Formulas regenerate weights at startup

### Links:
- 📦 [Formula Weights Repository](https://huggingface.co/arudradey/qwen-formula-engine)
- 🤖 [Base Model: Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct)

### Note
This Space requires ~2GB RAM to reconstruct and run the model. It uses CPU inference.