LM-Searcher: LLM-Powered Neural Architecture Search

LM-Searcher, developed by Ashenone3, is an 8 billion parameter language model specifically designed for cross-domain neural architecture search (NAS). Unlike traditional LLMs focused on text generation or understanding, LM-Searcher's core function is to act as a task-agnostic framework for exploring and optimizing neural network architectures.

Key Capabilities

• Automated Architecture Search: Utilizes an LLM to sample new architectural configurations within a defined search space.
• Task-Agnostic: Designed to be adaptable across various problem domains, allowing users to define custom reward functions for evaluation.
• Unified Numerical Encoding: Employs a method for encoding architectural designs that LLMs can process effectively.
• Scalable Deployment: Supports deployment via vllm for efficient inference during the search process.

Good For

• Researchers and Engineers: Ideal for those looking to automate the discovery of optimal neural network architectures.
• Experimentation: Facilitates rapid prototyping and evaluation of different architectural designs.
• Optimizing Performance: Can be integrated with custom evaluation metrics to find architectures that maximize specific performance indicators (e.g., accuracy, efficiency).

LM-Searcher provides a programmatic interface for defining search spaces and integrating custom reward functions, enabling a flexible approach to automated machine learning model design.