Sergei Vassilvitskii

We show that highly accurate LLMs can be learned from training sets consisting entirely of synthetic data and weakly curated data.