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.
1 comments
Someone showed embedding similarity is not enough to find data leakage.