They address this in the document, citing both an effort to eliminate test data from the training set, as well as the usage proprietary, ostensibly private test sets for evaluation.
Someone showed embedding similarity is not enough to find data leakage.