/

← All benchmarks

WildBench

instruction-following · real-user-tasks↗ Quelle

43Evidence

not yet measured

What it measures

Aufgaben aus echten Nutzerdialogen statt aus einem Aufgabenkatalog.

What it does not measure

Fachpruefungen, Codeausfuehrung.

How the score is composed

Currency50

Discrimination50

Headroom50

Contamination safety33

Gaming resistance55

Methodological rigor25

Reproducibility60

Details

Benchmaxxing risk: medium
Difficulty: medium
Public test set: yes
Human-verified: no
LLM-judged: yes
Skala: 0–100 WB-Score (höher besser)

Hand-maintained — no automatic adapter yet.

Best measured models

No results ingested yet.