Train a source classifier¶

A source classifier is a context-specific model trained from a site visit’s attached source data. It is for durable, governed workflows—not a standalone job shortcut.

Prerequisites¶

• a site visit with suitable source images and point labels
• a working labelset
• enough examples per target class
• a grouped mapping when the recipe requires one
• permission to edit the visit

Use the Source workspace¶

Open the site visit and select Source. The workspace separates:

• Overview — attachment and readiness
• Training — recipe and run controls
• Evaluation — metrics and comparisons
• Label mapping — detailed-to-grouped targets
• Reports & exports — available artifacts

Resolve the displayed Next step before training. A “mapping missing” state, for example, means the grouped target space is not yet defined.

Review recommendations¶

The Source Classifiers page is a review queue across accessible contexts.

A recommendation is evidence for model governance, not permission to skip evaluation. Compare class balance, validation results, failure cases, data lineage, and intended deployment scope before promotion.