Towards high-quality (maybe synthetic) datasets