How well generated molecules match the statistical properties of real molecules in a training dataset.