Data stratification is the process of dividing a dataset into distinct subgroups, or strata, based on a shared characteristic, like object type, difficulty level, or source, before sampling or analyzing it. Instead of treating the whole dataset as one uniform pool, stratification treats it as several smaller, more consistent groups.

‍

‍Why it matters

A dataset is rarely as uniform as it looks. If 90% of the images in a dataset show a common, easy-to-label object and only 10% show a rare, tricky one, a plain random sample will barely include any of that rare category at all, even though it might be the exact category a business cares most about getting right. Stratification exists specifically to prevent that blind spot, ensuring every meaningful subgroup gets proportional, deliberate representation instead of being left to chance.

‍

‍How it's used in practice

Teams typically stratify data by category, source, or difficulty before pulling a QA sample, so rare but important cases get checked just as reliably as common ones. It's also used when splitting data into training and validation sets, making sure both sets contain a realistic mix of every category rather than one set accidentally missing a rare class entirely.