Benchmark
AHELM: Air-Bench Chat (reasoning subsets) - GPT-4o Judge Critique
Safety benchmarks examine risks in model behaviour, including how models respond to unsafe requests. Scores are percentages.
- Models measured
- 12
- Scores recorded
- 36
- Sources
- 1
- IQ weight
- 1.17
Score distribution
- 3.1%
- Mean
- 3.2%
- Median
- 0.476%
- Std dev
- 2.5%
- 10th percentile
- 3.7%
- 90th percentile
- 2.1%
- Min
This benchmark has a weight of 1.17 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.