Improving language model behavior by training on a curated dataset