Tiny language models, pretrained entirely from scratch on a diverse and high quality mix of datasets.
Note V1