Upload models

What MLsuite accepts

MLsuite accepts trained models only as .mlbundle files exported with MLbundle. A bundle holds the model in its native, pickle-free format together with a manifest and a schema of its inputs and outputs. MLsuite never loads pickle or joblib files.

	Supported
Libraries	scikit-learn, XGBoost, CatBoost, ONNX
Tasks	Tabular classification and regression
Bundles	Written with the MLbundle version the runtime pins in backend/pyproject.toml (0.2.1 at the time of writing)
scikit-learn	The version the runtime pins (1.9.1 at the time of writing); a bundle from another version is refused

The bundle’s schema is used as it comes: export with training_data= to infer the form fields from the training frame, otherwise every feature becomes a number field. Fields map to the model’s own feature names, or to positions 0, 1, … for a model without names. One-hot columns are grouped into a single choice using the separator the bundle was exported with. Each model output gets its own report.

Trusted uploaders only

Model parsers and inference run unsandboxed in the runtime. Give upload permissions only to people you trust; see the security policy.

Upload

1. Open Models and choose New model (needs Create models).
2. Drop one or more .mlbundle files, or choose Add files. Each file is checked by the runtime first; a file that cannot be loaded is not added, and the reason appears as a message such as “Cannot load bundle: …”.
3. Check the name of each model. It defaults to the file name in title case and must be unique in the organization.
4. Choose Save on a card, or Save all to save every bundle in turn. When all are saved, MLsuite opens the models list.

Files up to 1000 MB pass the upload limits of a default installation; MLbundle itself refuses archives over 512 MiB. Models that a public bookmark or an API client runs are limited to 50 MB by default.

The models list

Search by name, file or algorithm; filter Active, Archived or All; sort by Latest updated, Name or Algorithm. Each row shows the kind and algorithm, who changed it last, and how many fields and reports its bundle schema has.

The model page

The header shows the kind, library, input and output counts, the file and its size. New schema starts a schema from this model. The tabs are:

Tab	Shows
Overview	The algorithm or pipeline steps, what it predicts (classes, or “A number”), bundle facts and feature importance when the model’s structure provides it.
Inputs	Every form field: its kind, which feature columns it feeds, the training type and whether it is required.
Outputs	Each report, its kind, output position, target and classes.
Parameters	The model’s hyperparameters, grouped by pipeline step and searchable.
Provenance	The stored file and its SHA-256, the bundle id and format, each member of the archive with its hash, the training environment (Python, system, package versions) and the author’s notes.

The training environment is what the exporter declared; MLsuite does not check it against the model.

Rename, duplicate, archive, delete

The actions of each row in the models list:

Action	Effect
Edit	Rename the model (needs Edit models).
Duplicate	Copy it under a new name (needs Create models).
Archive	Hide it from the active list. Archived models stay listed under Archived, open with an Archived badge, cannot be chosen for new schemas, and keep running in the schemas that already bind them. Archiving cannot be undone.
Delete	Remove it for good (needs Delete models). Refused while any schema snapshot or inference uses the model: archive it instead.

If someone else changed the model since you loaded the page, the action is refused with “Model changed since it was loaded; refresh and retry”. Reload and try again.