Reward model for RLHF trained on 3000 examples from Anthropic/hh-rlhf dataset.