“A hybrid approach combining RLAIF and RLHF can be effective, using a model's inherent ability to rank generations and then fine-tuning it with a small number of human-labeled examples, on the order of 50 to 100.”