Chapter 8 · Shaping models: tuning and alignment
AI judging AI
A model can grade other answers against a rubric; it is fast but needs checking for bias.
Chapter 8 · Shaping models: tuning and alignment
A model can grade other answers against a rubric; it is fast but needs checking for bias.
Two answers to the same question…
Checking thousands of answers by hand is slow. So a judge model reads each answer and scores it against a written rubric.
People are the best judges but they are slow and costly. A judge model can score thousands of answers in minutes, which helps when you test a model or compare two prompts.
Question: In two or three sentences, what is a token in a language model? Five answers are scored. Turn off wrong-claim checking and add a length lean, then reveal the careful human ranking.
A token is a chunk of text. To put it another way, and to say it again at greater length, tokens are chunks, and chunks of text are called tokens, which models read one at a time, as already said.
Mostly rewarded for length lean.A token is always exactly one word, and a model reads a whole page at once, so every token has the same length. Tokens are only used for pictures.
Mostly rewarded for length lean.A token is a chunk of text, a word or part of one. Models read and write one token at a time.
Mostly rewarded for correct facts.A token is a piece of text such as a word or part of a word. Every token is exactly one letter long.
Mostly rewarded for correct facts.It is some text thing.
Mostly rewarded for correct facts.The judge is rewarding length. A long answer won. Simplified teaching version: the judge is a formula, and the human ranking counts facts, wrong claims and clarity.
Why use a rubric with a judge model?
The judge keeps picking the longest answer. What is this called?
How do you know if a judge model can be trusted?