Do the problems you highlighted still appear with higher quality training data?