Tabular foundation models were pre-trained on synthetic data because a large corpus of real-world tabular data was not available