Llama 2 reported that a small set of human-written exam...

The Llama 2 result is not a universal law; it was highly dependent on the quality of the open-source synthetic datasets available at that time, which were often low-diversity and generated by weaker models. When AI-generated data is produced using a strong teacher, a diversity-enforced seed set, and automated quality filtering, it can consistently match or exceed small human-written corpora. Furthermore, the Llama 2 paper noted that RLHF helped recover much of the performance gap, suggesting that the human-annotation advantage was partially due to capturing preference signals rather than instruction quality alone.