It's basically impossible to overfit to the leaderboard in some Kaggles like Avazu where both the train and test are massive in terms of unique observations.