Balanced dataset

A dataset whose classes have similar numbers of examples, or follow a deliberately chosen balance.

Balance is a design choice, not a guarantee of fairness or accuracy. The useful distribution depends on the task and evaluation question. A balanced test set can help compare classes, but its overall score may not reflect performance where the real frequencies are different.

In practice

A balanced success and failure set can expose a weak success detector even when production data mostly contains successful trials.