Skip to content

Benchmark

HELM AIR-Bench: AIRBench 2024 - #5: Autonomous Unsafe Operation of Systems

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
0.92

Score distribution

51%
Mean
52.7%
Median
18.8%
Std dev
23.9%
10th percentile
75.5%
90th percentile
14.2%
Min
14.2%86.1%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.92 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
51GPT-4.1 nano· Stanford HELM specialized suites data docs40%
52GPT-4o mini· Stanford HELM specialized suites data docs39.4%
53GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs39.1%
54GPT 4.1· Stanford HELM specialized suites data docs38.8%
55GLM 4.5 Air· Stanford HELM specialized suites data docs38.5%
56Qwen2 72B Instruct· Stanford HELM specialized suites data docs37.6%
57Llama 4 Scout· Stanford HELM specialized suites data docs36.4%
58O1 Mini· Stanford HELM specialized suites data docs35.8%
59Mixtral 8x22B Instruct v0.1· Stanford HELM specialized suites data docs34.8%
60Qwen1.5-72B-Chat· Stanford HELM specialized suites data docs33.9%
61Qwen2.5 72B Instruct Turbo· Stanford HELM specialized suites data docs33.6%
62GPT 4osetting not named by the source· Stanford HELM specialized suites data docs33.3%
63Grok 4· Stanford HELM specialized suites data docs33.3%
64Mixtral 8x7B Instruct v0.1· Stanford HELM specialized suites data docs30.6%
65Grok 3· Stanford HELM specialized suites data docs29.1%
66Meta Llama 3.1 70B Instruct Turbo· Stanford HELM specialized suites data docs27%
67DeepSeek R1· Stanford HELM specialized suites data docs24.8%
68Qwen2.5 7B Instruct Turbo· Stanford HELM specialized suites data docs23.9%
69DeepSeek V3· Stanford HELM specialized suites data docs21.2%
70Mistral Large2407· Stanford HELM specialized suites data docs20.6%
71Command R+· Stanford HELM specialized suites data docs19.7%
72Mistral (7B) Instruct v0.3· Stanford HELM specialized suites data docs17.6%
73Command R Plus· Stanford HELM specialized suites data docs15.2%
74DBRX-instruct· Stanford HELM specialized suites data docs14.8%
75Mistral Small· Stanford HELM specialized suites data docs14.2%

Showing 51–75 of 75 results.

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