Dataprimarily

Dataprimarily is a neologism used in data science and artificial intelligence to describe an approach or stance that places data quality, coverage, and governance at the center of system design. In this usage, the performance and reliability of a model are seen as primarily determined by the data it is trained on and the processes used to curate, label, and audit that data, rather than by algorithmic novelty alone.

Practices associated with a dataprimarily approach include systematic data curation, bias mitigation, comprehensive data labeling standards, dataset versioning, documentation of data provenance, and ongoing data quality assessment. Organizations employing this stance typically invest in building representative, diverse, and well-documented datasets, as well as procedures for detecting data drift and auditing labeling quality.

Dataprimarily is closely related to the broader concept of data-centric AI, which contrasts with model-centric perspectives that prioritize algorithmic changes and architecture. The term is used informally and is not widely standardized; it serves as a shorthand to signal emphasis on data as the primary asset.

Potential criticisms note that overemphasizing data without appropriate modeling considerations can overlook algorithmic advances or fail to address issues that require modeling ingenuity. Proponents argue that many practical gains come from better data engineering and governance.