Data curation determines which examples enter a dataset, how duplicates and errors are handled, what metadata is retained and whether important cases are adequately represented. High-quality curation can matter as much as raw dataset size because models learn from the examples and structure they receive.
Curation is an ongoing process rather than a one-time cleanup. Product failures can reveal missing cases such as exact dimensions, identity preservation, brand terminology or repeated edits, allowing later datasets and evaluations to target weaknesses found in real use.
ELI5
Data curation is the careful work of choosing and organizing the examples used to build or test an AI. It is similar to preparing a useful library by removing broken copies, adding labels and making sure important subjects are represented.
For example, a media model may need examples of the same person across several edits so it can learn identity consistency. If those examples are missing or badly labelled, adding more unrelated data may not fix the problem.
