Skip to content

Benchmark

AHELM: Safety - Mean win rate

Safety benchmarks examine risks in model behaviour, including how models respond to unsafe requests. Scores are percentages. Higher is better.

Models measured
12
Scores recorded
12
Sources
1
IQ weight
0.03

Score distribution

37.2%
Mean
37.5%
Median
24.7%
Std dev
6.3%
10th percentile
68.8%
90th percentile
0%
Min
0%78.1%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.03 for chat models, measured over 9 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard