Skip to content

Trusted Python scenario files

A trusted Python file is useful for local/agent-driven experiments that should not yet be promoted into a discoverable library recipe.

Trusted execution

NexuML executes scenario files as Python code. Only run files you trust.

Contract

The file exposes scenario() -> ScenarioSpec:

from nexuml.core.types import DataSpec, LoaderSpec, ScenarioSpec, TrainingSpec
from nexuml.data.loaders.definitions import TorchLoader
from nexuml_library.data.synthetic import SyntheticDataset


def scenario() -> ScenarioSpec:
    return ScenarioSpec(
        name="experiment",
        data=DataSpec(
            source=SyntheticDataset(feature_shape=(64,), num_samples=1000),
            input_shapes={"features": [64]},
            loader=LoaderSpec(backend=TorchLoader()),
        ),
        pipeline=...,
        training=TrainingSpec(max_epochs=5),
    )

Run it directly:

nexuml train --scenario-file experiment.py

Save provenance

nexuml train \
  --scenario-file experiment.py \
  --artifact-dir artifacts/exp-001

The artifact directory records a source snapshot/hash and run provenance so an exploratory file can be tied to the resulting experiment.

Tuning metadata

Trusted scenario files can also expose the tuning metadata documented in Tuning file reference, including SEARCH_SPACE, TUNING_SPEC, and an optional structural build(**params) factory.

nexuml tune --scenario-file experiment.py --n-trials 20

When to promote it

Once a recipe is stable and should be reused by colleagues, move it into an importable library, decorate it with @scenario, and expose the package through nexuml.libraries. See Register a library.