I've been surprised that evaluation dataset leakage into training dataset is not discussed more, given that LLM creators often don't reveal their training data