Intermediate · 08

How we test language models

Benchmarks, leaderboards, human votes and AI judges: how anyone decides one model is better than another, and why it's harder than it looks.

7 min read · Tap underlined words for a definition

For a spam filter, testing is easy: count right and wrong answers on held-out emails. For a that writes essays and code, "right" is fuzzy. The field uses several kinds of test, each with blind spots. A test suite for models is usually called an eval.

Benchmarks

A fixed set of questions with known answers: maths problems, multiple-choice exams, coding tasks with unit tests. A benchmark gives one comparable number.

Weakness: once a is famous, its questions leak onto the web and into . A model can score well by memory, not skill. This is called contamination.

Human preference votes

Show people two anonymous answers to the same and ask which is better. Thousands of votes produce a ranking, like chess ratings.

Weakness: people favour confident, long, nicely formatted answers, even when a shorter one is more correct.

Quick check

Why can a model score highly on a famous benchmark without really having the skill?

AI as the judge

A strong model grades another model's answers against a rubric. It's cheap and fast, so teams can test thousands of cases on every change.

Weakness: judges have biases too, such as preferring answers that sound like themselves. Teams spot-check a sample by hand.

Your own evals beat leaderboards

If you build with AI, collect 50 to 100 real examples from your use case with what a good answer looks like. Run every model or prompt change against them. That tells you more than any public ranking.

Run itA tiny eval harness

# Pretend model: answers arithmetic questions, but badly at big numbers.
def model(q):
    a, b = map(int, q.split(" + "))
    return a + b if a < 50 else a + b + 1

tests = [("2 + 2", 4), ("10 + 7", 17), ("40 + 3", 43), ("75 + 5", 80), ("99 + 1", 100)]

passed = 0
for q, expected in tests:
    got = model(q)
    ok = got == expected
    passed += ok
    print(f"{'PASS' if ok else 'FAIL'}  {q} = {got}  (expected {expected})")

print(f"score: {passed}/{len(tests)}")

Edit anything, then run. Python runs in your browser.

Add your own test cases. Notice how a score alone hides where the model fails; reading the failures is where you learn.

Quick check

You're building a customer-support bot. What's the most useful way to compare two models?

Before you drift off

Benchmarks give comparable numbers but can be contaminated.
Human votes and AI judges measure quality, each with their own biases.
For real products, a small set of your own examples is the best eval.

Terms in this lesson

Go deeper, free and open

HELM: Holistic Evaluation of Language Models · Stanford CRFM
LMArena · LMArena · Human-preference leaderboard
The LLM Evaluation Guidebook · Hugging Face · Free

That's the lesson

How we test language models · 7 min

Finish & continue · Up nextSeeing and hearing: multimodal modelsIntermediate · 6 min

Was this lesson clear?

Spotted a mistake? Suggest an edit

Previous: Embeddings: meaning as maps