Concepts

MLsuite has a small vocabulary. Each term below names something the application stores and shows, and each one keeps a link to the others so that any result can be traced back to the exact model and form that produced it.

The objects

Organization : The workspace that owns models, schemas, bookmarks, inferences, reviews, plugins, API clients and members. Everything in the app belongs to the organization chosen in the switcher. Every account gets a personal organization when it is created.

Model : A trained model uploaded as an .mlbundle file. MLsuite stores the file, its SHA-256 and what the bundle declares: library, estimator, features, outputs, parameters and training environment.

Schema : The contract between people and one or more models: which fields the form asks for, how each field maps to a model feature, and which reports show the results. A schema is a named repository of snapshots.

Snapshot : An immutable, numbered version of a schema (v1, v2, …). It holds the form JSON and its bindings, the models it runs. Snapshots are never edited; changing a schema publishes a new one.

Change : Editable work on a schema, based on one snapshot. Publishing a change creates the next snapshot.

Bookmark : A named pointer to one snapshot, such as production. You run bookmarks, not schemas. A bookmark can be moved to a newer snapshot, published as a public page and given to API clients.

Inference : One saved run of a bookmark: the inputs, every bound model’s result, the snapshot it ran on, who ran it and from where (the workspace, a public page or the API). The API calls it a prediction run.

Review : A request for organization members to give feedback on a set of inferences of one snapshot, before a date. Each reviewer answers each inference.

Feedback : A reviewer’s answers about a result: the class or value they consider right for each built-in output, and the questions of the report’s feedback questionnaire.

Plugin : A TypeScript file that adds a field kind or report kind to the organization’s forms.

API client : A piece of outside software that runs the bookmarks it was given, with its own keys.

The loop

1. Upload a model. Its bundle carries the fields and reports that a schema starts from.
2. Create a schema from one or more models. Its first snapshot, v1, is published immediately.
3. Bookmark a snapshot. The bookmark appears under Predict.
4. Run the bookmark and save the runs you want to keep. Each saved run is an inference.
5. Ask reviewers to give feedback on a selection of inferences.
6. Filter the inferences, look at the answers in the table, and export them as CSV.

Improving the form means publishing a new snapshot and moving the bookmark to it. Earlier inferences keep pointing at the snapshot they ran on, so their inputs, outputs and feedback always keep their original meaning.

Who can do what

What a member sees depends on the permissions of their role in the organization: for example, View models shows Models, Schemas, Predict and Inferences, Run predictions lets them save runs, and Review opens the Review inbox. The built-in roles and every permission are listed in Roles and permissions.

A superadmin is a platform-wide account that can act in every organization and opens the administration pages.