Sweeps

Preview

The Sweeps tab is in early access. List and detail views work end-to-end, but sweep creation is API-only (Vantage SDK). See troubleshooting for the known limitations.

A sweep is a hyperparameter-search experiment that runs many trials, each trial is one training job with one specific parameter combination. Sweeps track the metric you're optimizing across all trials and surface the best one.

Sweep concepts

• Algorithms: the search strategy (Bayesian, grid, random) that determines how trials are generated.
• Presets: pre-defined configurations for common search patterns.
• Trials: individual parameter combinations within a sweep, each running as a training job.

Next steps

• Algorithms: the supported search strategies
• Presets: pre-configured sweep templates
• Trials: individual parameter combinations