Right. Training data might not encompass runs of translations exceeding this amount of characters so wouldn't be apparent for anything resembling existing training data.