Don't we also find with natural languages that focusing on training data from only a single language doesn't actually result in better writing in the target language, either?