Like PHI — textbooks are all you need. You can create entirely synthetic yet high quality training data with a strong model (the generated textbooks) and make very small models like PHI.