What if you could generate a full training dataset from 1 sentence and 0 seed examples? This new paper from @adaption_ai proposes Invent-a-Dataset, a specialized system that turns a dataset description directly into training data, rather than repeatedly prompting a general-purpose LLM It's able to beat frontier LLM prompting on both quality and diversity, with its diversity advantage growing to 37% at 20K samples Published exclusively on…