Supported models
mlbundle supports four libraries through closed adapters, plus scikit-learn pipelines that end in XGBoost or CatBoost. There is no conversion between formats and no generic fallback: a model outside an adapter’s boundary fails on export, on the training machine.
| Library | Extra | Stored as | Bundle member | Structure | Class labels |
|---|---|---|---|---|---|
| scikit-learn | sklearn |
skops, closed allowlist | model.skops |
Trees, ensembles, linear, importances; null for other families |
Native classes_ |
| XGBoost | xgboost |
Native UBJSON | model.ubj |
Complete JSON tree dumps | Supplied labels bound to a native copy, or encoded indices |
| CatBoost | catboost |
Native CBM | model.cbm |
Complete symmetric trees and borders | Native classes_ |
| ONNX | onnx |
Your file, byte for byte | model.onnx |
null |
Graph labels, or count-checked declarations |
| Pipelines | sklearn + native |
skops prefix + native leaf | model.skops + model.ubj/model.cbm |
Leaf structure with the transformed boundary | From the final estimator |
Every adapter owns recognition, export preparation, serialization, loading, claim and structure extraction, and prediction. The same extraction code runs on export and on load, so the format is defined once.
Shared limits
Section titled “Shared limits”- XGBoost, CatBoost and ONNX models have exactly one output. scikit-learn supports up to 1,024.
- At most 100,000 features and 100,000 classes per output.
- Classifiers must produce probabilities; there is no decision-only classifier support.
Explicitly out of scope
Section titled “Explicitly out of scope”mlbundle does not support:
- other libraries, including LightGBM and PyTorch;
- converting scikit-learn models to ONNX, or any other format conversion;
- per-model environments and bundle signing;
- pickle, joblib or cloudpickle, even as an option.
The normative mappings are in Adapters and structure, Prediction state, CatBoost mapping, ONNX mapping and Composite pipelines.