Yep, came here to say this. The big thing about the results here that might not be obvious to someone not in AI is that the models being trained in this paper are very many orders of magnitude smaller than the LLMs we've all heard so much about recently, and they're also being trained on specific tasks instead of general language modeling.

So I'm not expecting this will find its way into a LLaMA near me any time soon, but maybe this is an interesting result for people working in the specific domains represented in the evaluations.