Schema reference

A snapshot’s document is an MLform form schema with two lists, fields and reports, in which every mappedTo names the model it feeds. This is the document you edit in a change.

{
  "fields": [
    {
      "kind": "number",
      "label": "Age",
      "description": "Applicant age in years.",
      "step": 1,
      "required": true,
      "mappedTo": { "Demo Loan Tree": "age" }
    },
    {
      "kind": "onehot-category",
      "label": "Employment",
      "required": true,
      "options": [
        { "label": "Salaried", "value": "salaried", "mappedTo": { "Demo Loan Tree": "employment__salaried" } },
        { "label": "Self-employed", "value": "self_employed", "mappedTo": { "Demo Loan Tree": "employment__self_employed" } },
        { "label": "Unemployed", "value": "unemployed", "mappedTo": { "Demo Loan Tree": "employment__unemployed" } }
      ]
    }
  ],
  "reports": [
    {
      "kind": "classifier",
      "label": "Decision",
      "labels": ["approve", "decline"],
      "showClassProbabilities": true,
      "mappedTo": { "Demo Loan Tree": 0 },
      "feedbackQuestionnaire": { "steps": [] }
    }
  ]
}

Mapping to models

mappedTo is an object keyed by the model name of each binding, so one field can feed several models:

On a field, the value is the model’s feature name, or its position (0, 1, …) for a model that takes features by position.
On a onehot-category option, it is the one-hot column that option sets to 1; the other options’ columns are sent as 0.
On a report, it is the position of the model output the report shows.

The model’s id also works as a key. A snapshot must keep a field for every feature its models need; a model that misses a feature refuses the run.

Field kinds

Kind	Asks for
`number`	A number; `min`, `max`, `step`, `unit`.
`text`, `long-text`	Free text.
`boolean`	True or false; `trueLabel`, `falseLabel`.
`category`	One option from a list.
`single-choice`, `multi-choice`	One or several options shown at once.
`rating`	A score from `min` to `max`.
`date`	A date.
`onehot-category`	One option, sent as one 0/1 column per option.
`mapped-category`	One option that writes a preset of values into other fields.
`series`	A table of rows, sent as a list.

Every field and report may have an id. An explicit id is used exactly as written and cannot contain whitespace; without one, MLform derives it from the label in kebab case. Conditions refer to fields by these ids.

Every field also takes the shared options: label, description, required, defaultValue, and hidden, disabled and readOnly, each with a …When condition on other fields. The MLform field reference documents every option and conditions the …When syntax.

Kinds added by the organization’s plugins are accepted too.

Report kinds

Kind	Shows	Options
`classifier`	The predicted class and, with `showClassProbabilities`, a bar per class.	`labels`: the display name of each class, in the model’s class order.
`regressor`	The predicted value.	`unit` and the MLform report options.

Plugins can add report kinds. Classifier and regressor reports get an assessment step in reviews; every report can carry review questions in feedbackQuestionnaire, described in Feedback questionnaires.

What the bundle generates

Without edits, v1 holds what the bundle’s schema carried: a number, boolean, text, date, category, series or onehot-category field per feature, labelled with the feature name, and one report per model output, labelled “Predicted class” or “Predicted value”. When a schema combines several models, fields with the same label and kind are merged and each report is labelled “<report> · <model>”.