There isn't enough training data though, is there? The "secret sauce" of LLMs is the vast amount of training data available + the compute to process it all.
replyI think you could probably feed a copy of a toki pona grammar book to a big model, and have it produce ‘infinite’ training data
replyThis is essentially a distillation on the bigger model; you'd wind up surfacing a lot of artifacts from the host model, amplifying them in the same way repeated photocopying introduces errors.
replyhttps://dailyai.com/2025/05/create-a-replica-of-this-image-d...
There are not enough samples in that book to generate new "infinite" data.
reply