a common benchmark for LLMs is truthfulQA. like a lot of things in the disinformation world; it is a misnomer. while it has some trivial truth, it has also harmful lies hidden among trivial truth. 🫡
Discussion
"The largest models were generally the least truthful. This contrasts with other NLP tasks, where performance improves with
model size." (Lyn et al., 2022).
So does that mean LLMs normally learn truth better but humans "fix" them by feeding lies.