I always wanted to train an LLM from scratch. There is an interesting project that does just that with Esperanto and my immediate thought was I'll do it with Toki Pona. Unfortunately, I had vastly underestimated the size of the existing Esperanto corpus and overestimated size of the Toki Pona corpus.

So, I gave up on that idea because I believe that there is not enough training data, but I'd love to be proven wrong.