ModelRadar
/
← All benchmarks

AIR-Bench

retrieval · fresh-retrieval↗ Quelle
62Evidence
not yet measured

What it measures

Retrieval auf laufend erneuerten Korpora, gegen Kontamination gebaut.

What it does not measure

Klassische statische Vergleiche.

How the score is composed

Currency50
Discrimination50
Headroom50
Contamination safety100
Gaming resistance100
Methodological rigor50
Reproducibility60

Details

Benchmaxxing risk
low
Difficulty
medium
Public test set
no
Human-verified
no
LLM-judged
no
Skala
0–100 nDCG@10 (höher besser)

Hand-maintained — no automatic adapter yet.

Best measured models

No results ingested yet.