> Policy tuning is one way to remove it, but there's some research on applying p... | Hacker News

Hacker Newsnew | past | comments | ask | show | jobs | submit

		SrslyJosh on May 24, 2024 \| parent \| context \| favorite \| on: Google says no "African countries beginning with K... > Policy tuning is one way to remove it, but there's some research on applying policies like "this text is a lie" during pretraining instead of trusting all of it equally. Who's gonna vet all the training material?

astrange on May 24, 2024 [–]

The previous model does.

https://arxiv.org/abs/2309.00267

That's during fine-tuning though. I know I read one about applying it during pretraining, maybe this one?

https://arxiv.org/abs/2302.08582

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact