Is there any chance that this is because training corpus largely consists of documents shorter than the advertised context windows?