When training data includes examples from evaluation benchmarks, artificially inflating model performance metrics.