Dataset curation is the ongoing work of selecting, organizing, and maintaining the data that goes into a machine learning project, deciding what gets included, what gets excluded, and how the dataset evolves over time. It's a broader, more strategic task than labeling individual items, closer to editing a collection than producing one piece of it.

‍

‍Why it matters

‍ Not all data is worth collecting or keeping, duplicate items, low quality examples, or data that doesn't actually represent the real conditions a model will face can quietly drag down model performance even if every individual label is technically correct. Curation is what catches this at the dataset level rather than the item level, since a perfectly labeled dataset can still be a bad dataset if it's unbalanced, redundant, or missing entire categories a model will need to handle.

‍

‍What it involves in practice

Curators typically review a dataset for duplicate or near-duplicate items, check whether categories are proportionally represented, and remove data that's mislabeled, corrupted, or simply irrelevant to the task. This is usually an ongoing process rather than a one-time cleanup, since new data keeps arriving and a dataset that was well curated a year ago can drift out of balance as the real world it's supposed to represent changes.