Credo: Dataset

Make the data fit its intended use and establish what its construction supports. A schema-valid collection is not necessarily a suitable dataset; a reviewed sample is not proof that every record is correct.

Establish intended use

Recover the target task, intended population, unit of observation, source material, existing dataset, and permitted data handling. Identify the dataset decision: build a first version, revise a known problem, assess labels, inspect a split, or judge whether a claim is supported.

Inspect available records and representative examples before prescribing categories. Record inclusion and exclusion rules only where they change selection or interpretation. Prefer existing identifiers, schemas, storage, and versioning. A review returns findings; construction or revision requires the corresponding request.

Construct or inspect the data

Read construction and review for the relevant sampling, annotation, splitting, or lineage decision. Keep the normal work concrete:

• Relate sampling and coverage to the target population. Convenience sources and synthetic examples do not establish representative coverage.
• Define ambiguous labels using task-specific examples and counterexamples. Preserve disagreements rather than forcing uncertain items into a clean label.
• Separate independent label judgment from model-generated suggestions when independence matters. A generated label is a candidate until the required review establishes it.
• Choose splits by the unit of intended generalization, including related entities, time, documents, or source groups where needed. Check exact and meaningful near-duplication across those boundaries.
• Track transformations and exclusions enough to recover the dataset version and explain a consequential difference. Keep originals under the project's access and retention rules.