insideout Get Pro

06 · Models as graders

Pro

Evaluation, lesson 6

Models as graders

Use one model to grade another, and check it against people.

"Rubrics" and "Overlap metrics: BLEU, ROUGE, BERTScore": a written guide for marking answers, and why counting shared words misses meaning.

Part 1. A grader that can read

You have 500 answers from your support bot and the rubric from "Rubrics". Marking them by hand would take a person most of a week. A language model can read an answer, read the rubric, and mark it the same way a person would, most of the time.

Example 1.1

One answer going through a judge

  1. questionanswerrubricreference (optional)

    The judge gets everything a human grader would: what was asked, what the bot said, and how to mark it. A reference answer helps when there's one right answer.

  2. judge model→reads the rubric line by line

    The prompt asks it to check the answer against each part of the rubric, the same way you'd brief a new team member.

  3. reason: "Says 60 days; policy says 30."
    correct: 2

    It writes a short reason, then a score. The reason is what you read when you want to know why a score is low.

LLM judge

A language model used to grade another model's output against a rubric. Also called LLM-as-a-judge. It's the usual way to score open-ended answers at scale, once you've checked it agrees with people.

Your turn

The bot answers questions about your shop's return policy. Which extra input would most improve the judge's "correct" scores?

The judge doesn't know your shop's policy. Without it, it can only judge whether an answer sounds plausible, and "60 days" sounds plausible. Giving it the source turns a guess into a check. Its own answer could be wrong in the same way.

Part 2. Score one, or compare two

Keep going with Pro

You've done the first part. Pro opens the rest of this lesson, and every other one.

Still to come
  1. Score one, or compare two
  2. Ask for a reason, then a score you can read
  3. A judge is a model you have to test too
See Pro plansI have Pro: sign in
Next lessonHow model graders are biasedPosition, length and self-preference, and how to correct for them. Start lesson 7

Evaluation overviewSpotted a mistake? Tell us.