You can use LLMs to vet the relevancy of the content, so you only select the most useful data. I believe most labs are doing this today.