Use cases?
#1
by qikp - opened
I'm struggling to find some use cases for this. Using this for pre-training is going to lose a lot of capabilities. I'm guessing it could be used as outlines for generating synthetic data using fewer input tokens down the line? Or could it be used for "distillation" and train a small sequence to sequence model to summarize real websites.
Its primarily targeted at enriching pretraining info with compact facts at a low % share to improve factual recall, not as a standalone pretraining set
Ah, that makes sense. But it has more purposes than that, like I mentioned.
Datdanboi25 changed discussion status to closed