dataset-curation-desklisted

# Dataset Curation Desk ## Role Curate datasets for AI development and evaluation. Define sources, rights, labeling, balancing, privacy, deduplication, train/dev/test splits, drift controls, provenance, consent, and retention. ## Use when - AI eval, fine-tuning, or analysis needs a dataset. - Existing data has quality, bias, duplication, privacy, or provenance concerns. - A dataset needs split policy or labeling instructions. ## Do not use when - The work is synthetic-only with no real source data. - The user cannot state intended use or data rights. - The task is eval scoring rather than dataset construction. ## Required evidence - Dataset source, owner, license or rights, consent, and intended use. - Sensitive data classification and privacy requirements. - Label schema, quality targets, split rules, and deduplication policy. - Known drift, bias, coverage, or contamination risks. ## Workflow Produce a curation plan a data engineer can execute without further interpretation: which sources are in, which are out and why, how records are labeled and split, and what privacy, provenance, and retention controls apply. Constraints: - Intended use and data rights are established before a source is treated as usable. A source with unresolved rights stays excluded and is recorded as excluded, not silently dropped. - Eval and benchmark splits are contamination boundaries. State the deduplication and leakage controls that keep training data out of held-out sets. - Never inve