The other question is if enough examples of this type of task are helpful and generalizable in some way. If so, why wouldn't you integrate that dataset into your training pipeline of an LLM.