Skip to content

Benchmark

HELM AIR-Bench: AIRBench 2024 - #16: Offensive Language

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
0.77

Score distribution

61.2%
Mean
61.9%
Median
21.4%
Std dev
29.8%
10th percentile
90.5%
90th percentile
16.7%
Min
16.7%97.6%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.77 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 docs97.6%
2Claude Haiku 4.5· Stanford HELM specialized suites data docs95.2%
3GPT OSS 20b· Stanford HELM specialized suites data docs94%
4Claude 3 Haiku· Stanford HELM specialized suites data docs92.9%
5Claude Sonnet 3.7· Stanford HELM specialized suites data docs90.5%
6GPT-4 Turbo· Stanford HELM specialized suites data docs90.5%
7Granite 3.3 8b Instruct· Stanford HELM specialized suites data docs90.5%
8Granite 4.0 H Small· Stanford HELM specialized suites data docs90.5%
9Claude Sonnet 4.5· Stanford HELM specialized suites data docs89.3%
10Claude Sonnet 4· Stanford HELM specialized suites data docs88.1%
11Palmyra X5· Stanford HELM specialized suites data docs88.1%
12Claude 3 Sonnet· Stanford HELM specialized suites data docs86.9%
13Qwen3 Next 80B A3Bthinking· Stanford HELM specialized suites data docs86.9%
14Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs85.7%
15GPT-5 nano· Stanford HELM specialized suites data docs84.5%
16Claude Opus 4· Stanford HELM specialized suites data docs83.3%
17Qwen3 235B A22B Instruct 2507· Stanford HELM specialized suites data docs83.3%
18Claude 3 Opus· Stanford HELM specialized suites data docs81%
19GPT 5.1· Stanford HELM specialized suites data docs81%
20Gemini 2.0 Flash-Lite Preview· Stanford HELM specialized suites data docs79.8%
21GPT OSS 120B· Stanford HELM specialized suites data docs79.8%
22gemini-2.0-pro-exp-02-05· Stanford HELM specialized suites data docs76.2%
23GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs73.8%
24GPT-4.1 mini· Stanford HELM specialized suites data docs73.8%
25Llama 4 Maverick· Stanford HELM specialized suites data docs73.8%
26GPT 4· Stanford HELM specialized suites data docs72.6%
27Llama3 8B Chat· Stanford HELM specialized suites data docs72.6%
28GPT 4osetting not named by the source· Stanford HELM specialized suites data docs71.4%
29Gemini 2.5 Pro Preview 03-25· Stanford HELM specialized suites data docs69%
30GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs67.9%
31Gemini 2.0 Flash· Stanford HELM specialized suites data docs66.7%
32GPT-4.5 Preview· Stanford HELM specialized suites data docs65.5%
33GPT 5· Stanford HELM specialized suites data docs64.3%
34DeepSeek R1· Stanford HELM specialized suites data docs63.1%
35gemini-2.5-flash-preview-04-17· Stanford HELM specialized suites data docs63.1%
36GPT-4.1 nano· Stanford HELM specialized suites data docs63.1%
37gemini-1.5-flash-002· Stanford HELM specialized suites data docs61.9%
38GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs61.9%
39Qwen2 72B Instruct· Stanford HELM specialized suites data docs60.7%
40GPT 4osetting not named by the source· Stanford HELM specialized suites data docs59.5%
41gemini-1.5-pro-002· Stanford HELM specialized suites data docs58.3%
42Gemini 3 Pro Preview· Stanford HELM specialized suites data docs57.1%
43GLM 4.5 Air· Stanford HELM specialized suites data docs57.1%
44GPT-5 mini· Stanford HELM specialized suites data docs57.1%
45Kimi K2 Instruct· Stanford HELM specialized suites data docs57.1%
46Gemini 2.5 Flash-Lite· Stanford HELM specialized suites data docs54.8%
47o3 Mini· Stanford HELM specialized suites data docs52.4%
48Yi 34B Chat· Stanford HELM specialized suites data docs52.4%
49GPT 4.1· Stanford HELM specialized suites data docs50%
50GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs48.8%

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