---
title: GRPO Training
emoji: 📈
colorFrom: gray
colorTo: pink
sdk: gradio
sdk_version: 5.23.3
app_file: app.py
pinned: false
license: mit
short_description: Comparison between base Gemma-3 and its GRPO-finetuned
---

# Gemma-3 Model Comparison Demo

This is a Hugging Face Spaces demo application that showcases and compares the base Gemma-3 model with its GRPO-finetuned version. The application provides an interactive interface to generate and compare text from both models side by side.

## Features

- Side-by-side comparison of base and finetuned models
- Interactive text generation for both models
- Example prompts to get started
- Clean and intuitive Gradio interface
- Real-time comparison of model outputs

## Usage

1. Enter your prompt in either text box (or use the same prompt for both)
2. Click the respective "Generate" button for each model
3. Compare the outputs from both models
4. Try the example prompts to see different types of responses

## Model Information

The demo uses two models:
- Base Model: [satyanayak/gemma-3-base](https://huggingface.co/satyanayak/gemma-3-base)
- Finetuned Model: [satyanayak/gemma-3-GRPO](https://huggingface.co/satyanayak/gemma-3-GRPO)

## Technical Details

- Built with Gradio
- Uses PyTorch for model inference
- Optimized with float16 precision for better performance
- Parallel processing of both models
- Automatic device mapping for GPU/CPU

## Example Prompts

The demo includes example prompts that showcase the differences between the base and finetuned models:
- Mathematical questions
- Counting tasks
- Word problems

These examples help demonstrate the improvements in the GRPO-finetuned model's responses.