Datasets with human or AI feedback. Useful for training reward models or applying techniques like DPO.