Quick Start

Quick Start is the Quick Start page for neuropt, a tool for LLM-guided machine learning optimization. It describes a workflow in which a training function is run with different configurations, and an LLM reads the resulting training curves and metrics to design the next experiment.

The page shows the main pieces of that workflow. A user writes a training function that receives a config, trains a model, and returns results such as a required score plus per-epoch train and validation losses and any extra metrics such as accuracy. It also shows how a search space is defined with tuples for ranges and lists for categorical choices, with neuropt inferring sampling strategy from the parameter name and value type. The description notes that returning more information makes the LLM better able to reason about why a configuration worked, including spotting overfitting and underfitting.

Quick Start also documents two ways to run searches. One is a command-line interface with the command neuropt run train.py, with options such as --backend claude, -n 50, --log results.jsonl, and --device cuda. The other is a Python API using ArchSearch, where a search is created from a train function and search space, then run with a maximum number of evaluations and queried for the best configuration and best score. A separate example shows ArchSearch.from_model, where neuropt is given a model and figures out what to tune. The page says it works with PyTorch models, XGBoost, LightGBM, Random Forest, and any sklearn estimator.

Installation is shown with pip install "neuropt[llm]" and an API key for Anthropic is set with ANTHROPIC_API_KEY. The options table says --backend claude uses Claude and is the default if the API key is set, while --backend none enables random search without an API key. The page also lists CLI Reference and Python API sections, and its title and metadata identify neuropt as LLM-guided ML optimization.