LIMA: Less Is More for Alignment

There have been other approaches to make instruction tuning easier. This one proceeds largely on self generated synthetic data:

https://arxiv.org/abs/2212.10560

does it mean that industry specific data are no longer moats?