This model is trained from the char towards left.
To hard to distinguish the two Chars. So only one is trained.