---

name: multitask_mixture_split description: Construct prompted multitask training mixtures while enforcing held-out dataset boundaries for zero-shot evaluation.

Multitask Mixture and Held-Out Split

Use this skill after prompt rendering when you need an auditable P3/T0-style source-task mixture and held-out evaluation set. Do not use it to render templates or train a model.

Inputs

• Rendered prompted records with dataset, task family, template, source, and target metadata.
• A list of held-out dataset ids or task families.

Outputs

• train, eval, and diagnostics dictionaries showing record counts and template coverage.

Workflow

1. Partition records by dataset id and optional held-out task family.
2. Ensure no held-out dataset id appears in the training split.
3. Count templates per dataset to document prompt diversity.
4. Return JSON-serializable diagnostics for source-boundary review.

Validation

Run the included tests or Distiller skill validation with --run-tests.

Limitations

The skill implements deterministic split construction, not large-scale sampling weights from the full T0 training pipeline.