LIMA stands for less is more for alignment. These large language models roughly are trained in two stages. The first is this unsupervised pre-training from raw text. And the second part is large-scale instruction tuning and reinforcement learning to better align to end tasks and use of preferences.