# hyperopt — Hyperparameter tuning with local training

hyperopt is an AutoML service that suggests which model configuration to train next. You train and evaluate models on your computer, report validation losses, and receive tuning suggestions through the Tuning Client Python SDK.

## What stays on your computer?

Training datasets, model weights, and training execution stay local. The service receives configuration descriptions, model indices, validation losses, and progress information needed for tuning. Account and session records support the service; demos can also send optional usage events.

## Start an experiment

The SDK and optional CPU demos require Python 3.11 or newer. Install the package `tuning-client`; import it as `tuning_client`. Create and verify an account before using the service. Local reference schedulers can run without an account.

Install `tuning-client[demo,mlflow]>=0.2.2` for the demo and optional local dashboard. When the service demo offers to start MLflow, answer y. The dashboard shows the selected model, validation-loss curves and training allocations, with expandable hyperparameters. It stays available after training. Use `--mlflow-local` in scripts or `--no-mlflow` to skip it.


- [Install and run the breast-cancer demo](/docs/quickstart)
- [Understand observations, ranking, and data handling](/docs/how-it-works)
- [Inspect benchmarks and downloadable results](/benchmarks)
- [Read free-beta limits and future pricing](/pricing)
- [Get the SDK on PyPI](https://pypi.org/project/tuning-client/)

## Demo scope

The included demo compares twenty configurations with a tuning cap of 100 validation observations, versus 200 observations to fully train all twenty to the demo's cap. Afterwards, it finishes all models for evaluation. This is a tuning-stage observation budget, not a guarantee of half the total runtime, cost, or equal model quality.

Classification, regression, and small language and vision/language teaching datasets are included. Results depend on the dataset, model configurations, and training budget. The SDK and demo code were generated with ChatGPT and human reviewed; evaluate them for your use case.

## Contact

Email [hello@hyperopt.dev](mailto:hello@hyperopt.dev) or use the [feedback form](/#contact). Email is optional in the form; include it if you would like a reply.
