Chapter 6: LLM applications
Golden sets and LLM-as-judge
Open answers have no single right string, so a plain text match cannot grade them, and humans are slow. Interviewers often ask how you would show that a new prompt is better. A strong answer has three parts: a small human-scored test set, an LLM judge with a clear scale, and checks that the judge is fair.
What you will learn
- Build a golden set that covers key slices and edge cases
- Write a rubric-based or pairwise judge prompt
- Measure judge-human agreement and position bias
Key terms
- golden set
- a small, fixed set of answers that humans scored
- LLM judge
- an LLM that grades answers, often against a rubric
- Cohen's kappa ($\kappa$)
- agreement with humans after lucky matches are taken out; 1 is perfect
Prerequisites
Take RAG vs fine-tuning vs long context first. It comes right before this lesson in the course.
Where this lesson picks up: You chose between RAG, fine-tuning and long context: RAG for large facts that change, fine-tuning for behavior and format, long context for whole documents. Now you have a new prompt, and you must show that it is better.
What comes next
Next, you split a RAG answer's quality in two: did it fetch the right chunks, and does it stay true to them?
Next lesson: RAG and hallucination metrics
More lessons in this part
Take the lesson
Short cards, one idea at a time, with quick checks as you go.