|
|
--- |
|
|
license: mit |
|
|
title: GEPA Prompt Tuner |
|
|
emoji: 🏢 |
|
|
sdk: gradio |
|
|
sdk_version: 5.38.2 |
|
|
--- |
|
|
title: GEPA Prompt Optimizer emoji: 🧬🤖 colorFrom: blue colorTo: green sdk: gradio sdk_version: 4.31.0 python_version: 3.10 app_file: app.py pinned: false license: mit |
|
|
GEPA Prompt Optimizer for Hugging Face Models |
|
|
This Space is a functional implementation of the GEPA (Genetic-Pareto) framework, as described in the paper "GEPA: REFLECTIVE PROMPT EVOLUTION CAN OUTPERFORM REINFORCEMENT LEARNING". |
|
|
|
|
|
It allows you to automatically optimize a prompt for a target model (like google/gemma-2b-it) by having it "learn" from feedback on a small training set. |
|
|
|
|
|
🧬 How It Works |
|
|
The application uses an evolutionary approach to refine a "seed" prompt over several iterations: |
|
|
|
|
|
Selection: It selects a promising prompt from its current pool of candidates using a Pareto-based strategy, which favors diversity. |
|
|
|
|
|
Rollout: It runs the selected prompt on a task using your target Hugging Face model (e.g., Gemma). |
|
|
|
|
|
Reflection: It uses a powerful "reflector" model (Google's Gemini 1.5 Flash) to analyze the prompt's performance, the output, and detailed feedback. |
|
|
|
|
|
Mutation: The reflector model proposes a new, improved prompt designed to fix the observed failures. |
|
|
|
|
|
Evaluation: The new prompt is evaluated, and if it shows improvement, it is added to the candidate pool. |
|
|
|
|
|
This cycle repeats until the defined "rollout budget" is exhausted, leaving you with the best-performing prompt. |
|
|
|
|
|
🚀 How to Use |
|
|
Provide API Keys: |
|
|
|
|
|
Hugging Face API Token: Enter your Hugging Face token. This is used to run inference on the target model you want to optimize for. |
|
|
|
|
|
Google Gemini API Key: Enter your Gemini API key. This is required for the powerful "reflection" step. You can get a key from Google AI Studio. |
|
|
|
|
|
Configure the Optimization: |
|
|
|
|
|
Target Model ID: The Hugging Face model you want to create a prompt for (e.g., google/gemma-2b-it). |
|
|
|
|
|
Initial Seed Prompt: The starting prompt. Your goal is to improve this! |
|
|
|
|
|
Training Data: A small JSON dataset with input fields and evaluation criteria (e.g., expected_keywords). You must adapt this for your specific task. |
|
|
|
|
|
Budget: The total number of times the target model will be called. Higher budgets allow for more refinement but take longer. |
|
|
|
|
|
Start Optimization: Click the button and watch the logs to see the evolutionary process in action! The best prompt found will update in real-time. |
