OpenELM is an open-source library by CarperAI that enables evolutionary search with large language models (LLMs) for both code and natural language generation. It targets researchers and developers looking to explore novel AI capabilities through evolutionary algorithms, offering flexibility for local or API-based LLM integration and diverse computational profiles.

How It Works

OpenELM integrates LLMs with quality-diversity (QD) algorithms like MAP-Elites and a simple genetic algorithm. It supports prompt-based mutation, specialized diff models for code, and crossover with LLMs. LLMs are instantiated via Langchain, allowing broad compatibility with local HuggingFace models or API-based services. For high-throughput scenarios, it offers optional Nvidia Triton Inference Server support.

Quick Start & Requirements

• Install: pip install openelm
• For Sodarace environment: pip install swig then pip install openelm[sodaracer]
• See pyproject.toml for further install options.
• Official docs: https://github.com/CarperAI/OpenELM

Highlighted Details

• Supports MAP-Elites, CVT-MAP-Elites, Deep Grid MAP-Elites, and a genetic algorithm baseline.
• Integrates with LLMs via Langchain, supporting local and API-based models.
• Includes a secure code execution sandbox with gVisor for code generation domains.
• Provides example environments for robotics (Sodarace), image generation, programming puzzles, and prompt evolution.

Maintenance & Community

• Community channel: #openelm on the CarperAI Discord.

Licensing & Compatibility

• License: MIT.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

The project is actively developed, with some environments and features residing in experimental branches, indicating potential for ongoing changes and API instability.