I didn't quite get why RL was used instead of just collecting same data from labelers and either prompt-tuning, p-tuning or fine-tuning on that.