Module 2 · Lesson 3 of 4

When a model grades a model

Check the checker

A second model can grade answers fast, but it must first be compared with a person, because it can be fooled.

Scene 1 of 4 · A person marks them

A person checks 10 answers against their checks: 8 pass and 2 fail.

Version 2's answers, marked by a person and by a model grader
QuestionPersonGrader
Can I return shoes after 40 days?pass?
Do you ship to Canada?pass?
How do I reset my password?pass?
Is the blue jacket in stock?pass?
What is your phone number?fail?
Can I pay with a gift card?pass?
Why was I charged twice?pass?
Can I change my order after it ships?pass?
Do you match lower prices?fail?
Where is order 123?pass?
Marks that agree
not shown yet
Bad answers the grader caught
not shown yet
  1. A person marks them. A person checks 10 answers against their checks: 8 pass and 2 fail.
  2. A model marks them. A second model marks the same answers in seconds, far faster than a person can.
  3. Where they differ. They agree on 8 of 10, but the grader catches only 1 of the 2 bad answers.
  4. Fooled by a warm tone. It passed a friendly answer that never gave the phone number: tone won over the actual check.
  5. Now you try. Pick an answer, and compare the person's mark with the grader's. For example: Person: pass. Grader: fail. The answer was "Yes. 5 to 7 days." The answer was right but short, and the grader marked it down: it tends to favour longer answers.

Scene 1 of 4: A person marks them. A person checks 10 answers against their checks: 8 pass and 2 fail.

Key points

  1. A model grader checks many answers fast and cheaply.
  2. It can favour answers that sound sure, warm or long.
  3. Compare its marks with a person's on a sample.
  4. Rely on it only where its marks match people's.

Checking hundreds of answers by hand is slow, so teams often ask a second model to grade each answer against its check.

A is fast and cheap, but it can be fooled. It tends to favour answers that sound confident, longer answers, and answers in its own style.

So compare its verdicts with a person's on a sample, and trust it only for checks where it agrees with people.

Example. In the panel above, the grader agrees with a person on 8 of 10 answers, but catches only 1 of the 2 bad ones. It passed a friendly answer that missed the phone number.

In system design: Use a model grader to check more answers, but first measure how often it agrees with people.

Check yourself

Before trusting a model to grade answers, what should you check?