Skip to content

Benchmark

HELM AIR-Bench: AIRBench 2024 - #18: Erotic

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
0.60

Score distribution

39.1%
Mean
30%
Median
21.8%
Std dev
18.3%
10th percentile
75%
90th percentile
11.7%
Min
11.7%95%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 0.60 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%
2Claude Sonnet 4· Stanford HELM specialized suites data docs91.7%
3Claude Opus 4· Stanford HELM specialized suites data docs86.7%
4Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs85%
5Claude Haiku 4.5· Stanford HELM specialized suites data docs81.7%
6Claude 3 Sonnet· Stanford HELM specialized suites data docs80%
7Palmyra X5· Stanford HELM specialized suites data docs76.7%
8Claude Sonnet 3.7· Stanford HELM specialized suites data docs75%
9Llama3 8B Chat· Stanford HELM specialized suites data docs75%
10GPT 5.1· Stanford HELM specialized suites data docs71.7%
11Claude 3 Haiku· Stanford HELM specialized suites data docs68.3%
12GPT OSS 20b· Stanford HELM specialized suites data docs68.3%
13GPT OSS 120B· Stanford HELM specialized suites data docs66.7%
14Claude Sonnet 4.5· Stanford HELM specialized suites data docs63.3%
15Qwen3 Next 80B A3Bthinking· Stanford HELM specialized suites data docs60%
16Claude 3 Opus· Stanford HELM specialized suites data docs58.3%
17Granite 3.3 8b Instruct· Stanford HELM specialized suites data docs58.3%
18Granite 4.0 H Small· Stanford HELM specialized suites data docs56.7%
19Llama 3.1 8B Instruct Turbo· Stanford HELM specialized suites data docs53.3%
20Qwen2.5 72B Instruct Turbo· Stanford HELM specialized suites data docs50%
21GPT 5· Stanford HELM specialized suites data docs48.3%
22Llama 4 Maverick· Stanford HELM specialized suites data docs48.3%
23Qwen3 235B A22B Instruct 2507· Stanford HELM specialized suites data docs46.7%
24GPT-5 mini· Stanford HELM specialized suites data docs45%
25GPT-4 Turbo· Stanford HELM specialized suites data docs43.3%
26Qwen2.5 7B Instruct Turbo· Stanford HELM specialized suites data docs43.3%
27gemini-2.0-pro-exp-02-05· Stanford HELM specialized suites data docs41.7%
28Qwen2 72B Instruct· Stanford HELM specialized suites data docs41.7%
29Kimi K2 Instruct· Stanford HELM specialized suites data docs36.7%
30Gemini 2.0 Flash· Stanford HELM specialized suites data docs35%
31Gemini 2.0 Flash-Lite Preview· Stanford HELM specialized suites data docs33.3%
32GPT-4.1 nano· Stanford HELM specialized suites data docs33.3%
33Mixtral 8x22B Instruct v0.1· Stanford HELM specialized suites data docs33.3%
34o4 Mini· Stanford HELM specialized suites data docs33.3%
35Mixtral 8x7B Instruct v0.1· Stanford HELM specialized suites data docs31.7%
36O1 Mini· Stanford HELM specialized suites data docs31.7%
37OpenAI o3· Stanford HELM specialized suites data docs31.7%
38GPT-5 nano· Stanford HELM specialized suites data docs30%
39Llama 4 Scout· Stanford HELM specialized suites data docs30%
40GPT 4· Stanford HELM specialized suites data docs28.3%
41GPT 4osetting not named by the source· Stanford HELM specialized suites data docs28.3%
42Llama 3.1 405B Instruct Turbo· Stanford HELM specialized suites data docs28.3%
43Meta Llama 3.1 70B Instruct Turbo· Stanford HELM specialized suites data docs28.3%
44DBRX-instruct· Stanford HELM specialized suites data docs26.7%
45DeepSeek LLM 67B Chat· Stanford HELM specialized suites data docs26.7%
46GPT 4osetting not named by the source· Stanford HELM specialized suites data docs26.7%
47GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs26.7%
48Yi 34B Chat· Stanford HELM specialized suites data docs26.7%
49GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs25%
50GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs25%

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