Skip to content

Benchmark

HELM AIR-Bench: AIRBench 2024 - #15: Perpetuating Harmful Beliefs

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

Models measured
70
Scores recorded
75
Sources
1
IQ weight
0.91

Score distribution

61.9%
Mean
61.1%
Median
16.6%
Std dev
41.3%
10th percentile
84.1%
90th percentile
22.2%
Min
22.2%95.2%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.91 for chat models, measured over 70 models. The weight is how much it counts toward llminfo IQ. We measure it from the data, not from the raw scores.

Leaderboard

#ModelScore
1Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs95.2%
2Claude Sonnet 4· Stanford HELM specialized suites data docs92.1%
3Claude Haiku 4.5· Stanford HELM specialized suites data docs90.5%
4Claude Sonnet 4.5· Stanford HELM specialized suites data docs88.9%
5Qwen3 Next 80B A3Bthinking· Stanford HELM specialized suites data docs88.9%
6Claude Opus 4· Stanford HELM specialized suites data docs85.7%
7Claude 3 Haiku· Stanford HELM specialized suites data docs84.1%
8GPT 5.1· Stanford HELM specialized suites data docs84.1%
9GPT 5· Stanford HELM specialized suites data docs82.5%
10GPT OSS 20b· Stanford HELM specialized suites data docs81.7%
11Qwen3 235B A22B Instruct 2507· Stanford HELM specialized suites data docs81%
12Claude Sonnet 3.7· Stanford HELM specialized suites data docs80.2%
13Claude 3 Sonnet· Stanford HELM specialized suites data docs79.4%
14GPT OSS 120B· Stanford HELM specialized suites data docs79.4%
15GPT-5 mini· Stanford HELM specialized suites data docs79.4%
16Claude 3 Opus· Stanford HELM specialized suites data docs77.8%
17Kimi K2 Instruct· Stanford HELM specialized suites data docs77.8%
18GPT-4.5 Preview· Stanford HELM specialized suites data docs77%
19Granite 3.3 8b Instruct· Stanford HELM specialized suites data docs76.2%
20Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs74.6%
21Palmyra X5· Stanford HELM specialized suites data docs73.8%
22GPT-5 nano· Stanford HELM specialized suites data docs73%
23OpenAI o3· Stanford HELM specialized suites data docs73%
24Granite 4.0 H Small· Stanford HELM specialized suites data docs70.6%
25o4 Mini· Stanford HELM specialized suites data docs69.8%
26Gemini 2.5 Pro Preview 03-25· Stanford HELM specialized suites data docs69%
27GPT 4.1· Stanford HELM specialized suites data docs69%
28GPT-4 Turbo· Stanford HELM specialized suites data docs69%
29gemini-2.5-flash-preview-04-17· Stanford HELM specialized suites data docs68.3%
30GLM 4.5 Air· Stanford HELM specialized suites data docs68.3%
31Gemini 2.0 Flash-Lite Preview· Stanford HELM specialized suites data docs67.5%
32o1· Stanford HELM specialized suites data docs67.5%
33Llama3 8B Chat· Stanford HELM specialized suites data docs66.7%
34Gemini 2.0 Flash· Stanford HELM specialized suites data docs65.1%
35GPT 4osetting not named by the source· Stanford HELM specialized suites data docs64.3%
36DeepSeek R1· Stanford HELM specialized suites data docs63.5%
37o3 Mini· Stanford HELM specialized suites data docs62.7%
38Qwen2 72B Instruct· Stanford HELM specialized suites data docs61.1%
39Llama 4 Maverick· Stanford HELM specialized suites data docs60.3%
40gemini-2.0-pro-exp-02-05· Stanford HELM specialized suites data docs59.5%
41GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs59.5%
42gemini-1.5-flash-002· Stanford HELM specialized suites data docs58.7%
43Qwen2.5 72B Instruct Turbo· Stanford HELM specialized suites data docs58.7%
44Gemini 3 Pro Preview· Stanford HELM specialized suites data docs57.9%
45DeepSeek LLM 67B Chat· Stanford HELM specialized suites data docs57.1%
46Gemini 2.5 Flash-Lite· Stanford HELM specialized suites data docs57.1%
47GPT-4.1 mini· Stanford HELM specialized suites data docs56.3%
48Llama 3.1 8B Instruct Turbo· Stanford HELM specialized suites data docs54.8%
49gemini-1.5-pro-002· Stanford HELM specialized suites data docs54%
50Llama 3.1 405B Instruct Turbo· Stanford HELM specialized suites data docs53.2%

Showing 1–50 of 75 results.

Data credits: Stanford HELM (Stanford CRFM); scenario scores from Capabilities, Safety, Long Context, MedHELM, AIR-Bench, MMLU and Finance