undefined | Better HN

0 pointszone4117mo ago0 comments

GPT-5 set a new record on my Confabulations on Provided Texts benchmark: https://github.com/lechmazur/confabulations/

0 comments

For how much I’ve seen it pushed that this model has lower hallucination rates, it’s quite odd that every actual test I’ve seen says the opposite.

alkyon7mo ago

Maybe its training set included this repo?

j / k navigate · click thread line to collapse

0 pointszone4117mo ago0 comments

GPT-5 set a new record on my Confabulations on Provided Texts benchmark: https://github.com/lechmazur/confabulations/

For how much I’ve seen it pushed that this model has lower hallucination rates, it’s quite odd that every actual test I’ve seen says the opposite.

alkyon7mo ago

Maybe its training set included this repo?

j / k navigate · click thread line to collapse