Element <ml-train>

Purpose: Train and publish a configured machine-learning artifact from project data.

Why: Use it to train and publish a configured model artifact from project-owned data.

Example

<ml-train name="customer_churn" source="analyticsDb"/>

Decision guide

Business value: Produces a versioned tabular model artifact for use cases that require learned distributions.

• Use when

  • A governed project explicitly requires model training from an owned dataset.

• Choose another approach when

  • Deterministic rules or registered generators can express the required data contract.

• Prerequisites

  • Provide an authorized source, privacy exclusions, resource bounds, and the ML test profile.

• Alternatives

  • Use deterministic generation when business rules define the required values. (See: <generate>)

Complete examples

Train a bounded tabular model and consume it through ml://

Use ml-train only when a named source dataset, privacy exclusions, and bounded training policy are explicit requirements of the model.

customer_id|segment|age
C-001|retail|31
C-002|business|48
C-003|retail|26
C-004|business|55
C-005|retail|39

<setup defaultSeparator="|" numProcess="1">
    <ml-train name="customer_model"
              source="data/customers.ent.csv"
              maxTrainingTime="1"/>
    <generate name="synthetic_customers"
              count="10"
              source="ml://customer_model"
              distribution="ordered"
              target="LogExporter">
        <key name="model_source" constant="customer_model"/>
    </generate>
</setup>

Rules and invalid combinations

W004 — Deprecated XML Attribute Ignored

<{element}> attribute '{attribute}' in descriptor '{descriptor}' is deprecated and ignored; execution continues. {migration}

Why: The descriptor uses a known legacy attribute that no longer controls execution.

Resolution: Follow the migration hint when updating the model. Removing the attribute is not required to run it.

Full rule

Allowed parents / Allowed children

Allowed parents: else, else-if, if, setup, while

Allowed children:

None

Attributes

Show all 35 attributes

analyzeDropColumns

Columns to drop during analyze stage.

optional; string; Default: null.

analyzeValueProtection

Enable analysis for value protection.

optional; boolean; Default: null.

batchSize

Batch size for training.

optional; integer; Default: null.

contextSource

Optional relational context source for two-table training.

optional; string; Default: null.

contextType

Type of the relational context.

optional; string; Default: null.

ctxPrimaryKey

Primary key column(s) in context; supports composite.

optional; string; Default: null.

device

Computation device selector passed to the ML engine.

optional; string; Default: null.

differentialPrivacy

Differential privacy configuration (JSON).

optional; string; Default: null.

encodingTypes

Per-column encoding overrides.

optional; string; Default: null; Values: AUTO, TABULAR_CATEGORICAL, TABULAR_NUMERIC_AUTO, TABULAR_NUMERIC_DISCRETE, TABULAR_NUMERIC_BINNED, TABULAR_NUMERIC_DIGIT, TABULAR_CHARACTER, TABULAR_DATETIME, TABULAR_DATETIME_RELATIVE, TABULAR_LAT_LONG, LANGUAGE_TEXT, LANGUAGE_CATEGORICAL, LANGUAGE_NUMERIC, LANGUAGE_DATETIME.

fairness

Fairness configuration (JSON).

optional; string; Default: null.

flexibleGeneration

Enable flexible generation mode.

optional; boolean; Default: null.

generationBatchSize

Batch size during generation.

optional; integer; Default: null.

gradientAccumulationSteps

Steps for gradient accumulation.

optional; integer; Default: null.

imputation

Imputation configuration (JSON).

optional; string; Default: null.

maxEpochs

Maximum number of training epochs.

optional; number; Default: null.

maxSequenceWindow

Maximum sequence window size.

optional; integer; Default: null.

maxTrainingTime

Maximum training time in minutes.

optional; string; Default: null.

mode

Deprecated: accepted with a warning and ignored. This setting does not control training or persistence.

optional; string; Default: null.

modelStateStrategy

Strategy for managing model state.

optional; string; Default: null; Values: reset, resume, reuse.

modelType

Model type/preset used by the engine.

optional; string; Default: null; Values: TABULAR, LANGUAGE.

name

Name of the ML model.

required; string.

rareCategoryReplacementMethod

Method for handling rare categories.

optional; string; Default: null; Values: constant, sample.

rebalancing

Class/feature rebalancing configuration (JSON).

optional; string; Default: null.

samplingTemperature

Sampling temperature used for generation (0-2).

optional; number; Default: null.

samplingTopP

Top-p (nucleus) sampling threshold (0-1).

optional; number; Default: null.

sensitiveColumns

Columns to drop for privacy (comma-separated).

optional; string; Default: null.

separator

Separator for generated data.

optional; string; Default: null.

source

Training data source: a project file or configured source identifier.

required; string.

splitPartitions

Number of partitions for dataset split.

optional; integer; Default: null.

targetColumn

Target column for supervised training.

optional; string; Default: null.

tgtContextKey

Join key(s) linking target to context; supports composite.

optional; string; Default: null.

tgtPrimaryKey

Primary key column(s) in target; supports composite.

optional; string; Default: null.

trainModel

Training model identifier or preset.

optional; string; Default: null.

trainValSplit

Train/validation split ratio (0-1).

optional; number; Default: null.

type

Type of data to train on.

optional; string; Default: null.