It didnt, just small enough, which is good, its way more useful to train small purposeful models than to use an entire frontier LLM.