“Rejection sampling, as used by Meta for Llama 2, involves using a reward model to select the best few of many generated responses and then using that curated set for further instruction tuning.”