We used a system of LLM critics to generate a high-quality dataset for this. More in the blog post linked in the answer above.