An AI governance lead is preparing a primer on training methods. Which definition best captures Reinforcement Learning from Human Feedback (RLHF)?

• AA training method that learns from labeled (input, correct-output) pairs only
• BA training method that finds clusters in unlabeled data without supervision
• CA training method where a reward model is fit to human preference comparisons and used to fine-tune a base language model to produce outputs humans prefer
• DA compression technique that quantizes model weights to lower precision