View only. Sign in to run this demo. You can also run it with your own model key, without an account.
LLM-as-Judge
Score extractor output, detect hallucinations, compare texts, and run a multi-model judge panel.
How to use the LLM-as-Judge
- 1Pick a modeFour ways to check an extractor: compare two texts, find invented fields, validate against a PDF, or ask a panel of judges.
- 2Start from a caseEach mode has ready-made cases that say what a person would expect the judge to find.
- 3Run the judgeA model reads the case and gives a verdict, a 1 to 5 score and the reason.
- 4Read the verdictSee whether it matches the expectation, which words or fields it is about, and what to correct.
- A judge is a model that decides whether an extracted value is right. It handles paraphrases, word forms and date formats that an exact comparison calls wrong, but it can be wrong too: read the reason, not only the verdict.
- The score reads: 1 is clearly wrong, 2 to 3 is suspect and worth a look, 4 to 5 looks correct. Several judges together (the panel) show how sure the verdict is.
Start from a case
1. The two texts
Compare an extractor's value with a reference. The judge handles paraphrasing, word forms and date formats, and needs no source document.
The name helps the judge: "quantity" is read as a number, "date" as a date.
No judgement yet
Pick a case or fill in the two texts, then click Judge equivalence.
How the score reads (Likert 1–5):
1wrongClearly wrong
2–3suspectSuspect: review before accepting
4–5okLooks correct
