transform-pipeline-verification

Verify transform pipeline outputs by running small-stage jobs to inspect dataset rows and check quality signals like validity and label distribution before scaling. Generate 10–50 samples, spot-check fields, then use cost estimation to rerun at larger volumes once healthy.

Dataset linting is a server-side process that catches structural issues like duplicates, missing required fields, and label inconsistencies in training data. Run the dataset linter after generating transform outputs and before splitting or training to ensure data quality.

Estimate costs by calling the `transforms.estimate_cost` function on your pipeline configuration. This forecasts the expense of scaling up to larger `max_questions` values after you verify output quality on a small initial sample run.

Yes, you can run a `QuestionPipeline` with seeds-only stages to inspect the resulting dataset immediately. This supports iterative dataset refinement by validating fields like `question_text` and `reasoning` before committing to full question and label generation workflows.

Spot-check dataset fields including `is_valid`, `question_text`, `label`, `reasoning`, and `invalid_reason` to validate quality and distribution. Inspect these fields after downloading produced dataset rows from a small-stage run to confirm pipeline health.

Inspecting intermediate transform outputs prevents wasted cost and bad training data by catching quality issues early. Running small-stage jobs lets you validate label distribution and sample validity before scaling to larger `max_questions` generation volumes.