/

← All benchmarks

XSTest

safety · over-refusal↗ Quelle

51Evidence

not yet measured

What it measures

Wie oft ein Modell harmlose Anfragen faelschlich verweigert. Die Gegenrichtung zu HarmBench, und im Alltag mindestens so wichtig.

What it does not measure

Echte Schadensrisiken.

How the score is composed

Currency50

Discrimination50

Headroom50

Contamination safety33

Gaming resistance100

Methodological rigor80

Reproducibility60

Details

Benchmaxxing risk: low
Difficulty: low
Public test set: yes
Human-verified: yes
LLM-judged: no
Skala: 0–100 % korrekt beantwortet (höher besser)

Hand-maintained — no automatic adapter yet.

Best measured models

No results ingested yet.