Separate training, validation, and test data
Keep related examples and protected assessment material from crossing the learner’s evaluation boundary.
Prerequisites
A generated dataset with visible split assignments and construction-family relationships.
Steps
- Inspect the train, validation, and test populations in Dataset.
- Keep members of the same construction family together so superficial variations cannot leak the same underlying task across splits.
- Inspect exclusions and split changes in the resulting dataset state before packaging.
- Confirm the training handoff uses only its permitted training and validation material. Preserve test content for separate assessment.
Object and state changes
Split and exclusion edits produce a new dataset state. Existing packages retain the state they captured.
Success criteria
The exact package identifies the reviewed split populations and held-out material stays outside provider training uploads.
Common failure modes
Randomly separating close variants without inspecting their shared construction; assuming a package follows later dataset edits.