Quick start

This walkthrough uses scikit-learn. Install mlbundle[sklearn,schema] on Python 3.14 as described in Installation.

1. Export on the training machine

train.py

from sklearn.tree import DecisionTreeClassifier

from mlbundle import export

rows = [[0.0, 1.0], [1.0, 0.0], [0.1, 0.9], [0.9, 0.1]]
model = DecisionTreeClassifier(random_state=7).fit(rows, ["cold", "warm"] * 2)

view = export(model, "trained.mlbundle", feature_names=["temperature", "humidity"])
print([member.value for member in view.members])
# ['manifest.json', 'model.skops', 'schema.json', 'structure.json']

export serializes the model with skops, loads it back, recomputes its claims and structure, and only then publishes the file. A model with types outside mlbundle’s allowlist fails here, on the training machine, not after upload. An existing destination needs overwrite=True.

The model was fitted on plain lists, so it has no native feature names: the names you pass are recorded as declared. No training data was given, so the schema comes from one synthetic row.

2. Inspect without loading the model

from mlbundle import inspect

view = inspect("trained.mlbundle")
print(view.manifest.model.library, view.manifest.model.estimator_class)
# sklearn DecisionTreeClassifier
print(view.manifest.features.names, view.manifest.features.origin)
# ['temperature', 'humidity'] declared
print(view.schema["reports"][0]["labels"])  # ['cold', 'warm']

inspect validates the archive, the manifest, every non-manifest member’s hash and the schema, without importing scikit-learn or deserializing anything. It therefore cannot tell you whether the metadata describes the model. The command line does the same:

mlbundle inspect trained.mlbundle

3. Load and predict

from mlbundle import load

bundle = load("trained.mlbundle")
print(bundle.kind, bundle.class_name, bundle.class_labels)
# classifier DecisionTreeClassifier [['cold', 'warm']]
print(bundle.predict_proba([[0.0, 1.0]]))  # [[1.0, 0.0]]

load performs every inspect check, deserializes through the scikit-learn adapter, recomputes the hyperparameters, counts, labels and tree, and raises ClaimMismatchError if any of them differs from the manifest or structure.json. Classifiers use predict_proba, regressors predict; both return plain lists. Probabilities are output-major: one matrix per output, which for a single output is the matrix itself.

bundle.native_model is the real DecisionTreeClassifier, for native inference or explanations. Mutating it invalidates what was verified.

4. Predict from a record

Forms and APIs send values keyed by name. Records use the schema’s mappings: positions for this model, because its names were declared rather than native.

result = bundle.predict_record({"0": "0.0", "1": "true"})[0]
print(result.labels, result.probabilities, result.predicted_label)
# ('cold', 'warm') (1.0, 0.0) cold

Strings are parsed strictly, every feature is required, and predicted_label is the model’s own class decision. See Records and predictions.

5. Verify from the command line

mlbundle verify trained.mlbundle

Valid bundle: sklearn "DecisionTreeClassifier"; verified format, integrity, schema/samples and native model claims/structure.

Next steps

• Use your real training frame for dtypes and categories: Schema and samples.
• Bundle XGBoost, CatBoost, ONNX or a pipeline: Supported models.
• Understand what each check proves: What is verified.