Training on bot data can be bad when it's ungrounded and basically hallucinations on hallucinations. Having LLMs help curate something grounded is generally re

Training on bot data can be bad when it's ungrounded and basically hallucinations on hallucinations. Having LLMs help curate something grounded is generally reasonable. Functionally, it's somewhat similar to how some training is using generated subtitles of videos for training video/text pairs; it's very feasible to also go and clean those up, even though it is bot data.