One of the target columns in y_train is made up of a constant value. It seems catboost has issues learning if all observations in the training data have the same outcome.