The last page picked one token at a time. Whether the sentence they add up to is any good is a separate question — every score below is real in the sense that someone measured something, so click the asterisk to see exactly what that something was, and what it wasn't.
Category names and scores are illustrative examples, not live results from any specific model — the point is the shape of the caveat, not the number.
So what
Page 8 showed nothing inside the model checks its own facts. Nothing inside it grades itself either — every number on this page comes from outside.
MODULE RECAP
Click an idea below, then click the card it matches.
Underneath every page in this series: attention connects words, parameters get tuned, tokens are the real unit, sampling has more knobs than one, and nothing scores itself.
Back to the Topic Atlas