Some classification problems do not have a balanced number of examples for each class label. For example, for a simple classification task, there might be 20 samples labelled frog, and 80 samples labelled airplane.
Stratification preserves this proportion of the each labels in each of the sets.
Stratified split is the correct choice when the data is skewed in terms of the labels, like the aforementioned example.
Only 13% of vision AI projects make it to production, with Hasty we boost that number to 100%.