And here's a great reader project who trained a tokenizer from scratch on Nepali:

→ View original post on X — @rasbt