CourseLarge Language Models: The Complete Course · Module 10: Evaluation · part 71 of 103
Part 71 · Module 10: Evaluation

Lesson 3: Human evaluation: two raters, one rubric, one number

18 min read·13 Oct 2026
Module 10 · Lesson 3 · About 45 minutes · Needs: Module 4 Lesson 2 (labels from written definitions); Module 10 Lesson 1 helps (deterministic checks)

The situation

It is Thursday 13 August 2026, week 11. Two days ago Ravi approved a German draft that promised Brandt Logistik a full refund it was not due. Joel stopped the money at the approval step, but the promise had been sent.

RAVI'S BATCH, 11 AUGUST: 40 DRAFTS, 25 MINUTES draft 31: the incident 37.5 seconds per draft
Ravi approved the incident draft as number 31 of 40 in a 25-minute batch: 37.5 seconds per draft (red marks the failure).

Maya asks: "If one of us can miss that, how far can we trust any of our ratings?" In week 10 Lena and Ravi both rated the same 20 drafts, blind. You will measure their agreement, settle their disagreements and write a labelled set. The running example is the draft for T-1001, Ana Duarte's duplicate charge of 288 USD: Lena rated it "edit", Ravi "send". All ratings and drafts here are added in this lesson and simulated.

The rest of this course is yours to keep

This course is bought on its own, once, and stays readable afterwards, including the parts added to it later.