Your trained model, schema and verified structure. One file.

MLbundle packages a trained model with its manifest, locally inferred input schema and extractable structure in one compressed .mlbundle file. The same library writes the bundle and validates it when loaded, preserving feature order, class labels and output meaning across the training/runtime boundary.

Use it to hand a model to MLSuite, inspect its metadata or build your own consumer around the validated native model.

Why bundle a model?

A model file alone leaves the next application to recover its inputs, labels, environment and supporting metadata. Separate descriptions can drift from the model they are supposed to describe.

MLbundle keeps them together and checks model-derived claims against the loaded model. Schema inference runs locally through MLschema; real training rows are included only when explicitly requested. Native formats preserve the supported model without conversion.

Installation

MLbundle 0.2.1 requires Python 3.14 for export with the schema extra. The installation command for this release is:

python -m pip install "mlbundle[sklearn,schema]==0.2.1"

To use a local checkout, run python -m pip install -e ".[sklearn,schema]" from its root. The base library requires Python >=3.14; schema inference uses mlschema >=0.3.0,<0.4 on Python 3.14 only.

Choose xgboost, catboost or onnx instead of sklearn, or install [all]. Export always needs schema. Metadata-only inspection needs only the base package; load and CLI verification need the matching model extra.

Quick start

Train a model, export its schema without including training rows, then inspect and load the bundle:

import pandas as pd
from sklearn.tree import DecisionTreeClassifier

from mlbundle import export, inspect, load

training = pd.DataFrame(
    {
        "temperature": [0.0, 1.0, 0.1, 0.9],
        "humidity": [1.0, 0.0, 0.9, 0.1],
        "target": ["cold", "warm", "cold", "warm"],
    }
)
model = DecisionTreeClassifier(random_state=7).fit(
    training.drop(columns="target"), training["target"]
)

export(
    model,
    "trained.mlbundle",
    training_data=training,
    target_names=["target"],
)

metadata = inspect("trained.mlbundle")  # No model deserialization.
print(metadata.manifest.features.names)  # ["temperature", "humidity"]

bundle = load("trained.mlbundle")  # Recomputes native claims and structure.
result = bundle.predict_record({"temperature": 0.0, "humidity": 1.0})[0]
print(result.predicted_label)  # cold
print(result.probabilities)  # (1.0, 0.0)

The target column is excluded from the input schema. Add sample_rows=N only when you intend to include the first N real feature rows. Without training_data=, export infers a default schema from synthetic values.

Existing destinations require overwrite=True. Native feature names take precedence; otherwise mappings use zero-based positions. load and inspect accept paths, bytes or streams. The native model remains available as bundle.native_model; mutating it invalidates the verified metadata.

Choose your operation

Operation	Purpose
export(model, destination, ...)	Validate a native round trip, infer the schema locally and write a bundle.
inspect(source)	Validate the container, manifest, hashes and declarations without loading a model.
load(source)	Load through a closed adapter and verify model-derived claims and structure.
bundle.predict_proba(data)	Classifier probabilities as plain lists.
bundle.predict(data)	Regressor predictions as plain lists.
bundle.predict_record(record)	Map one record through the schema and return frozen output-major results.
bundle.row_from_record(record)	Validate and order record values for native inference or explanations.

The CLI exposes metadata inspection and full verification:

mlbundle inspect trained.mlbundle
mlbundle inspect trained.mlbundle --json
mlbundle verify trained.mlbundle

Inspection does not prove that metadata describes the model. Full verification loads the model and compares its claims. Neither operation establishes authorship or training history.

Supported models

Library	Storage	Support boundary
scikit-learn	skops	Closed fitted classifier/regressor families and supported transformers.
XGBoost	Native UBJSON	Exact single-target tree wrappers and closed objectives.
CatBoost	Native CBM	Exact numeric, single-target symmetric classifiers/regressors.
ONNX	Original ONNX bytes	Self-contained tabular models with supported signatures; no conversion.
Composite pipelines	skops + UBJSON or CBM	Supported sklearn preprocessing with an XGBoost/CatBoost final estimator.

Custom types fail at export. Sklearn and CatBoost own their class labels; XGBoost can bind declared labels to a native copy. ONNX has no visualization structure, and all-sklearn Pipeline structure is null. The adapter contracts define the exact support limits.

Some notes

Package versions and wire-format versions are independent. Format 1 is not frozen; incompatible schema shapes require re-export. The security design has not had an independent review; do not accept bundles from untrusted parties.

The reader bounds ZIP/JSON data, verifies integrity and uses closed native loaders. Hashes detect corruption, not deliberate tampering. Native parsers and inference still need process isolation and application-owned resource limits. Read SECURITY.md for guarantees, limits and private reporting.

Training rows stay local by default, but feature names, categories, labels and learned structure can reveal information. Schemas and environment records remain author declarations. Model training, hosting, signing and conversion are outside the library's scope.

Documentation

The documentation site includes installation, adapter guides, record prediction and API reference. It renders the normative contracts from this repository unchanged:

• Bundle format and verification
• Schema, mappings and explicit samples
• Supported adapters and structure
• Sklearn state and trusted types
• Composite pipelines
• ONNX layouts and bounds

The playground provides editable examples and inspection of the generated .mlbundle file. It runs scikit-learn examples in the browser and inspects uploaded files without loading their models.

Contributing

The Python library uses uv and Python 3.14:

uv python install 3.14
uv sync --locked --all-extras
uv run pre-commit install
uv run ruff check .
uv run ruff format --check .
uv run pyright
uv run pytest
uv build

Tests use the installed package and enforce 80% branch-inclusive coverage. CI validates the built wheel on Windows and Linux. Read AGENTS.md for contribution rules and publishing for the automatic release workflow.

The landing, docs and playground use Vite+ and a pnpm workspace. With vp installed, run from the repository root:

vp install
vp run check
vp run build

Start a site with vp run dev:landing, vp run dev:docs or vp run dev:playground. Their local ports are 4340, 4341 and 10021 respectively. Each site also accepts vp run dev, vp run build and vp run preview from its directory. Astro sites run their framework CLI through vp exec; the playground uses Vite+ directly. vp run check runs frontend lint, landing formatting, all three sites' type checks, and the docs and playground unit tests. Runtime and browser tests live in playground/.

Actions runs Python and frontend checks on pushes and PRs to main and develop. After all checks pass on main, it deploys the three tested site builds to Cloudflare. Manual CI Pipeline runs on main use the same gates and can redeploy without publishing Python. A change to project.version on main automatically starts the PyPI workflow after its version and branch preflight. See website deployment and publishing.

License

MIT. The design takes ideas from MLflow's versioned manifest and native storage, not its code.