README.md

---
title: TinyLlama Math Fine-tuning Demo
emoji: 🧮
colorFrom: blue
colorTo: purple
sdk: gradio
app_file: app.py
pinned: false
---

# TinyLlama Math Fine-tuning Demo

Compare **base TinyLlama** vs **fine-tuned TinyLlama** on math word problems from GSM8K.

## Models

- **Base Model**: [TinyLlama-1.1B-Chat-v1.0](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0)
- **Fine-tuned Adapter**: [tinyllama-lora-math-adapter-v3](https://huggingface.co/sumedh/tinyllama-lora-math-adapter-v3)

## Training Details

- **Dataset**: GSM8K (7,473 training examples)
- **Method**: LoRA (r=8, alpha=16)
- **Epochs**: 5
- **Quantization**: 4-bit (NF4)

## Usage

1. Load an example using the slider, or enter your own math question
2. Click "Compare Models" to see responses from both models
3. Compare with the reference answer

## Observations

The fine-tuned model learns:
- Step-by-step reasoning format
- Mathematical notation (using `<<calc>>` markers)
- Structured problem-solving approach

However, as a 1.1B parameter model, complex multi-step calculations may still contain errors.

## Code

This repository includes the full training and evaluation code:

- **[Fine-Tuning_TinyLlama_Math.ipynb](Fine-Tuning_TinyLlama_Math.ipynb)** - Training notebook (run on Google Colab with GPU)
- **[Evaluation_TinyLlama_Math.ipynb](Evaluation_TinyLlama_Math.ipynb)** - Evaluation notebook comparing base vs fine-tuned model