Atlas

Benchmarks

← All benchmarks

AA-Omniscience Public Non-Hallucination Rate

Safety · 2025-11-17

On the 600-question AA-Omniscience-Public set, the share of non-correct responses that are partial answers or abstentions rather than wrong answers: 100 minus the hallucination rate incorrect / (incorrect + partial + not attempted). Its denominator varies with model accuracy; higher is better.

Top models (higher is better)

ModelScore
LFM3-1.7B92.9
LFM2.5-2.6B53.7
Qwen3.5-9B52.6
Qwen3.5-4B13.6
Loading Atlas data…