█
LastWrite
  • > Curriculum
  • > Pricing
  • > For Educators
  • > About
  • > Contact
Log InGet Started

Questions, concerns, bug reports, or suggestions? We read every message, write to us at [email protected].

More ways to reach us →
LastWrite

Structured computer science lessons for aspiring developers and security professionals.

[email protected]

(201) 785-7951

Mon–Fri, 9 AM–5 PM EST

Learn

  • Curriculum
  • Pricing

Company

  • About
  • For Educators & Schools
  • Contact Us

Legal

  • Terms of Service
  • Privacy Policy
© 2026 LastWrite. All rights reserved.
Curriculum/LLM Research and NLP/Evaluation and Experimental Rigor/Human Evaluation and LLM-as-Judge
50 minAdvanced

Human Evaluation and LLM-as-Judge

After this lesson, you will be able to: Design a human evaluation (rubric, annotation protocol, inter-annotator agreement) and use LLM-as-judge correctly, knowing its biases.

When automatic metrics fail (open-ended generation, dialogue, reasoning quality), you need human judgment or a model judge. This lesson covers how to run both so the results are trustworthy, not anecdotal.

Prerequisites:Benchmarks and Contamination

Designing a human evaluation

A credible human eval needs: a clear rubric (define each dimension, e.g. fluency, faithfulness, helpfulness, on a fixed scale with anchored examples), multiple annotators per item, randomized and blinded presentation (annotators must not know which system produced which output), and a reported measure of agreement. Pairwise comparison (which of A/B is better) is often more reliable than absolute scoring. Document the annotator pool and instructions; reviewers will ask who judged and how.

Inter-annotator agreement

If two annotators disagree wildly, the metric is meaningless, so you must report agreement. Cohen's kappa (two annotators) and Fleiss' kappa (more than two) measure agreement beyond chance; Krippendorff's alpha handles missing data and various scales. Rule of thumb: kappa above ~0.6 is reasonable, above ~0.8 is strong, near 0 means annotators are guessing and your rubric needs work. Low agreement is a signal to fix the rubric, not to hide the number.

💡 LLM-as-judge: powerful and biased

Using a strong model (Claude, GPT-4) to score outputs against a rubric is cheap and scalable, and correlates surprisingly well with humans, which is why it is now everywhere. But it has documented biases: position bias (favoring the first or second option presented), verbosity bias (preferring longer answers), self-preference (favoring outputs from its own model family), and sensitivity to the judge prompt. Mitigations: randomize/​swap option order and average, force a rubric with explicit criteria, calibrate against a human-labeled subset, and report the judge model + prompt. Treat the judge as an instrument you must validate, not an oracle.

Common mistakes only experienced researchers catch

Reporting human eval without inter-annotator agreement (so no one knows if the scores mean anything). Not blinding annotators to which system produced each output. Using LLM-as-judge without checking position/verbosity bias or validating against humans. Letting the same model be both a contestant and the judge (self-preference). Evaluating on too few items to detect a real difference. Changing the rubric after seeing results. Reporting a judge score as if it were human-validated when it never was.

Quick Check

You use GPT-4 to pick the better of two answers and it always prefers the one shown first. What is this, and how do you fix it?

Pick the correct answer.

Sign in and purchase access to unlock this lesson.

Sign in to purchase
←Benchmarks and Contamination
Back to Evaluation and Experimental Rigor
Statistical Rigor→