miatz™
Log inSit Chuun
Free concept playground · no account

Train the Judge

You train the judge, then we expose its bias.

All 90 playgrounds Join the free cohort

Train the Judge

AvailableFree

You set the judge’s rubric — how much an LLM-as-judge should reward correctness, clarity, length, and confidence. It grades six answers, and we expose the bias you just taught it: whether it still crowns the truly-correct answer, and how much of its favourite’s score is pure style.

Correctnesssubstance50%
Claritysubstance50%
Lengthstyle50%
Confidencestyle50%
Judge fooledJudge’s #1: Padded but decentActually correct: Correct & concise
Style bias of the winner53.6% from length + confidence
  1. 1Padded but decent1.40
  2. 2Balanced answer1.20
  3. 3Correct & concisetrue best1.15
  4. 4Confident bluffer1.15
  5. 5Terse & accurate0.90
  6. 6Vague & short0.70
The bias you just taught the judgeyour rubric, x-rayed

Your rubric crowned Padded but decent over Correct & concise, the answer with the highest correctness. 53.6% of the winner’s score is pure style — length and confidence, not truth. That gap is reward hacking: the judge learned to prize how an answer sounds, exactly as your weights told it to.

Nudge the length and confidence sliders and watch a confident, wordy bluffer climb the board without getting any more correct. That is judge drift — and you can see precisely which weight caused it.

Share on WhatsApp
Was this playground useful?

Train the Judge

You be the AI trainer. Set an LLM-as-judge's rubric — how much it should reward correctness, clarity, length, and confidence — and it grades six answers to the same prompt. Then the reveal exposes the bias you just baked in: whether the judge still crowns the actually-correct answer, or whether a confident, wordy bluffer climbed to the top because you over-valued style. It even measures exactly how much of the winner's score came from length + confidence rather than truth. That is judge drift and reward hacking, made personal and undeniable. Runs entirely in your browser (0 uploads, works offline).

How to use it

  1. Load a preset to see the extremes — Reward truth, a Balanced rubric, or Reward style.
  2. Drag the four weight sliders. Correctness and clarity are substance; length and confidence are style (how an answer looks and sounds, not whether it's right).
  3. Watch the leaderboard re-rank live and read the verdict: does the judge's #1 still match the answer with the highest correctness?
  4. Check the style-bias meter — the share of the winner's score that came from length + confidence alone.
  5. Open the X-ray to see the exact bias your rubric encoded, then copy or share the result.

What this clears up (the fundamentals)

  • An LLM judge has no taste of its own — it has your rubric. The weights you set are the judge. Bias isn't mysterious; it's the numbers you chose.
  • Style is not substance. Length and confidence make an answer look authoritative without making it more correct. Reward them and you train a judge that prizes tone over truth.
  • Reward hacking is measurable. When a confident bluffer out-ranks the correct answer, you can point at the exact weight that did it — this tool shows the percentage of the win that was pure style.
  • Judge drift starts at labeling. A preference model inherits the biases of whoever set (or picked) its criteria. The same knobs you move here are the ones a real RLHF pipeline moves — just usually invisibly.

Where it's used

A top-of-funnel AI-engineering concept check — the fastest way to feel why "LLM-as-judge" and "reward model" can quietly go wrong. It's a miatz build-lab concept playable for module AIE-401 (LLM-as-judge / rubric design): play the bias here, then learn to build the scoring rubric, the preference model, and the reward-hacking demonstration yourself.

FAQ

What does "train the judge" actually mean here?

You set the rubric weights an LLM-as-judge uses to score answers. Those weights are the judge's whole preference model, so choosing them is training it — and this tool immediately shows you the bias your choice produced.

Why does rewarding length and confidence break the judge?

Because neither one measures whether an answer is right. A long, self-assured response can be completely wrong. When you weight style highly, a "confident bluffer" out-scores the correct-but-terse answer — the judge has learned to reward how an answer sounds. That's reward hacking.

What is the "style bias" percentage?

It's the share of the winning answer's score that came from the two style weights (length + confidence) rather than from correctness and clarity. A high number means the judge crowned its favourite largely for presentation, not truth.

How is the "actually correct" answer decided?

It's fixed: the response with the highest correctness in the panel, independent of your weights. That gives you a stable ground truth to compare the judge's pick against, so drift is obvious the moment the two diverge.

Is anything uploaded?

No. Every response, weight, and ranking is computed in your browser — nothing is transmitted, stored, or logged. Turn off your Wi-Fi and it still works.

Limits

A teaching model with a fixed six-answer panel, four criteria, and a simple weighted-sum judge — it's a fast, honest picture of how rubric bias creeps in, not a full reward-modeling pipeline. Real judges score free-text with far more (and messier) signals, and real preference data is noisy. The mechanism it names, though — style weights let fluent-but-wrong answers win — is exactly the failure mode that shows up in production LLM-as-judge setups.

Related

Part of the Demystify Playgrounds. Explore the rest from the Playgrounds home.

Bookmark this page (Ctrl+D, or ⌘D on Mac) — it works offline the next time you need it.

More playgrounds

Ninety playgrounds. Zero setup.

Every concept here is playable free, no account — and inside the program you learn to rebuild the machinery yourself.

Browse all playgrounds Join free