GEPA Prompt Optimizer for Hugging Face Models

This Space implements the GEPA (Genetic-Pareto) framework to automatically optimize prompts for a target model (like Gemma) hosted on Hugging Face. It uses a powerful LLM (Gemini) for the "reflection" step to propose high-quality prompt improvements.

Important Notes:

Make sure you have accepted the model's license on Hugging Face
Some models require a Pro subscription for Inference API access
Start with a small budget (5-10) to test connectivity first

1. Configuration

Hugging Face API Token (Optional)

Leave blank if HF_TOKEN is set as a Space Secret.

Google Gemini API Key (Optional)

Leave blank if GEMINI_API_KEY is set as a Space Secret.

Target Model ID

The Hugging Face model to optimize for. Try 'microsoft/DialoGPT-medium' or 'gpt2' for free models.

Initial Seed Prompt

Training Data (JSON)

Optimization Budget (Total Rollouts)

5 50

2. Results

Best Prompt Found

Best Score

Optimization Log