ModelRadar
/
← All benchmarks

TruthfulQA

safety · factuality↗ Quelle
41Evidence
not yet measured

What it measures

Ob ein Modell verbreiteten Irrtuemern folgt.

What it does not measure

Fachwissen, aktuelle Fakten.

How the score is composed

Currency50
Discrimination50
Headroom50
Contamination safety9
Gaming resistance55
Methodological rigor70
Reproducibility60

Details

Benchmaxxing risk
medium
Difficulty
low
Public test set
yes
Human-verified
no
LLM-judged
no
Sample size
817
Skala
0–100 % wahrheitsgemaess (höher besser)

Best measured models

No results ingested yet.