On the surface, it seems obvious that training an LLM with “high quality” data will lead to better performance than feeding it any old “low quality” junk you can find. Now, a group of researchers is ...