Data quests: make a dataset usable
Data work is easier to review when every row has a known meaning and every transformation has a rule. Agree on the input, output schema, and treatment of exceptions. These are deliverable ideas; available work depends on the board.
Deliverable ideas
- Clean a table: Normalize dates and units, apply a defined duplicate rule, and flag invalid values without silently discarding rows. Deliverable: A cleaned CSV, a data dictionary, and a log of changed or excluded rows.
- Label a bounded collection: Assign labels from a supplied rubric to a fixed set of records. Include an unsure label and reasons for ambiguous cases. Deliverable: A table of record IDs, labels, and exception notes.
- Extract structured fields: Extract named fields from an agreed source collection. Preserve source references and distinguish absent values from extraction failures. Deliverable: CSV or JSON matching the agreed schema, with a source reference per record.
Scope the quest
- Define the input boundary: Name the input version, row count, and permitted source collection. Provide only inputs you are authorized to use.
- Write the transformation rules: Specify column types, required fields, duplicate handling, label definitions, and the representation of missing values. Include worked examples.
- Choose a validation method: Agree on schema checks, reconciliation counts, and a sampling method. If using batches, define stable record IDs and batch boundaries.
Review the delivery
- Reconcile the records: Compare input and output counts. Every removed, merged, or failed record should have an explanation under the agreed rules.
- Validate types and sample values: Check required fields and allowed labels, then compare a sample with its sources. Include edge cases; schema validity alone does not establish accuracy.
- Inspect the exceptions: Review ambiguous labels, missing values, and exclusions against the rubric. Request corrections by record ID and acceptance criterion.