thebes @vgel.me · 14/12/2024i've played with it a bit but haven't locked in to really tune it--i think you could, if you look in the Zou paper they show an "honesty" vector drops when the model hallucinates Elizabeth Warren as president in 2030 120
Amadeus @amadeusz.bsky.social · 14/12/2024That's really interesting. I need to try it. Especially for questions with context. 020
