One verifiable file for a trained model.

mlbundle puts the model in its native format, a manifest of model and training-environment metadata, a locally inferred input schema and the model's structure into one .mlbundle file. Before it hands you the model, the reader recomputes every model-derived claim and rejects disagreement.

Quick start What is verified Format 1 contract

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)
export(model, "trained.mlbundle", feature_names=["temperature", "humidity"])

from mlbundle import load

bundle = load("trained.mlbundle")  # verified before it is returned
print(bundle.predict_proba([[0.0, 1.0]]))  # [[1.0, 0.0]]
print(bundle.predict_record({"0": "0.0", "1": "true"})[0].predicted_label)  # cold

No pickle, joblib or cloudpickle is involved at any step. scikit-learn models are stored with skops against a closed allowlist, XGBoost as native UBJSON, CatBoost as native CBM, and ONNX files byte for byte. Loading rebuilds the model through mlbundle’s own adapter, extracts its claims and structure with the same code the writer used, and rejects the bundle when they disagree.

Find your path

Bundle a first model Export a classifier, inspect it, load it and predict. Control what goes in Feature names, labels, targets, metadata and overwrite. Keep data local What the schema carries, and why samples need sample_rows=N. Read a bundle inspect for metadata, load for a verified model. Predict from a form Records keyed by the schema, frozen per-output results. Check your model scikit-learn, XGBoost, CatBoost, ONNX and composite pipelines.

Why a bundle

Uploading a pickled model and a DataFrame leaves three gaps that mlbundle closes:

Gap	What mlbundle does
Loading a pickle can execute code.	Closed adapters use skops, native UBJSON/CBM or the original ONNX file. There is no pickle fallback.
The training versions are lost.	The manifest records Python, OS, architecture and installed package versions next to model and feature metadata.
Training data leaves the machine.	Schema inference runs locally. Rows are included only when you ask for them explicitly.

mlbundle is the training-side companion to MLSuite, whose runtime accepts only .mlbundle uploads. It is open source under the MIT license and part of eshkra. Read the source on GitHub.