Run scenarios¶
NexuML accepts different scenario sources at different boundaries. Use this page to choose the source; use the generated CLI reference for exact flags.
Supported sources¶
| Command | Registered scenario | Resolved YAML | Trusted Python file |
|---|---|---|---|
resolve |
yes | — | — |
build |
— | yes | — |
train |
yes | yes | yes |
export-dataset |
yes | yes | — |
tune |
yes | — | yes |
export |
yes | — | — |
smoke |
yes | — | — |
Only one scenario source may be supplied to a command.
Registered scenario¶
Libraries expose scenario recipes with @scenario:
nexuml registry list scenarios
nexuml resolve my-scenario
nexuml train my-scenario
Use this for stable, reusable experiment recipes.
Resolved YAML¶
resolve persists the scenario's typed definitions and surrounding specs:
nexuml resolve my-scenario -o configs/my-scenario.yaml
nexuml build configs/my-scenario.yaml
nexuml train -c configs/my-scenario.yaml
Use YAML when you want an inspectable/versionable frozen configuration rather than re-evaluating the scenario recipe.
Trusted Python file¶
For local experiments, a file can expose scenario() -> ScenarioSpec:
nexuml train --scenario-file experiment.py
nexuml tune --scenario-file experiment.py --n-trials 20
Use --artifact-dir when you want NexuML to snapshot the trusted file and provenance alongside the run.
Trusted execution
Scenario files are executed as Python code. Only run files you trust.
Checkpoint-only resume¶
train also accepts a Lightning trainer checkpoint without a separate scenario source. NexuML can recover the persisted scenario information from a compatible checkpoint. See Checkpoints.