data

Class Imbalance

A dataset condition where some labels appear far more often than others.

Class imbalance describes a dataset where the distribution of labels is skewed — one category might make up 90% of examples while another makes up under 1%. This is common and often unavoidable: real-world data on harmful content, rare defects, or unusual user intents is naturally scarce compared to common cases.

The practical consequence is that models trained on imbalanced data tend to perform well on frequent classes and poorly on rare ones, since there's simply less signal to learn from. Teams address this through targeted collection of underrepresented examples, hard negative mining, or reweighting during training — but the annotation side of the fix starts with schema coverage analysis to even identify which classes are underrepresented.

Class imbalance also complicates evaluation: raw accuracy on an imbalanced dataset can look high while performance on the rare-but-important class is actually poor, which is why precision and recall broken out per class matter more than an aggregate score.

What this means for trainers

If you notice you're rarely seeing examples of a category you know exists, that imbalance is worth surfacing — targeted sourcing of rare cases is often a separate, sometimes better-paid task.