Skip to content

Benchmark

HELM MMLU: Security Studies - EM

Scores are percentages. Higher is better.

Models measured
64
Scores recorded
68
Sources
1
IQ weight
1.86

Score distribution

77.5%
Mean
80%
Median
9.6%
Std dev
68.2%
10th percentile
86.1%
90th percentile
40.8%
Min
40.8%88.6%
The band covers the middle 80% of scores. The line marks the median.

This benchmark has a weight of 1.86 for chat models, measured over 64 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 Opus· Stanford HELM specialized suites data docs88.6%
2Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs88.2%
3Claude 3.5 Sonnetsetting not named by the source· Stanford HELM specialized suites data docs87.8%
4gemini-1.5-pro-001· Stanford HELM specialized suites data docs87.3%
5Claude 3 Sonnet· Stanford HELM specialized suites data docs86.5%
6Mistral Large2407· Stanford HELM specialized suites data docs86.5%
7mixtral-8x22b· Stanford HELM specialized suites data docs86.5%
8GPT 4· Stanford HELM specialized suites data docs86.1%
9gemini-1.5-pro-002· Stanford HELM specialized suites data docs85.7%
10Llama 3.1 405B Instruct Turbo· Stanford HELM specialized suites data docs85.7%
11Llama-3.2-90B-Vision-Instruct-Turbo· Stanford HELM specialized suites data docs85.3%
12Meta Llama 3.1 70B Instruct Turbo· Stanford HELM specialized suites data docs84.9%
13Nova Prov1· Stanford HELM specialized suites data docs84.9%
14Qwen2.5 72B Instruct Turbo· Stanford HELM specialized suites data docs84.9%
15Llama 3.3 70B Instruct Turbo· Stanford HELM specialized suites data docs84.5%
16GPT 4 1106 Preview· Stanford HELM specialized suites data docs84.1%
17DeepSeek V3· Stanford HELM specialized suites data docs83.7%
18GPT 4osetting not named by the source· Stanford HELM specialized suites data docs83.7%
19Qwen2 72B Instruct· Stanford HELM specialized suites data docs83.7%
20GPT 4osetting not named by the source· Stanford HELM specialized suites data docs83.3%
21Meta Llama 3 70b· Stanford HELM specialized suites data docs83.3%
22Yi-34B· Stanford HELM specialized suites data docs83.3%
23Phi-3 Medium 4K Instruct· Stanford HELM specialized suites data docs82.9%
24text-unicorn· Stanford HELM specialized suites data docs82.9%
25Mistral Largesetting not named by the source· Stanford HELM specialized suites data docs82.4%
26Qwen1.5-72B· Stanford HELM specialized suites data docs82.4%
27Claude 2.1· Stanford HELM specialized suites data docs81.2%
28Claude 3 Haiku· Stanford HELM specialized suites data docs80.8%
29gemini-1.5-flash-001· Stanford HELM specialized suites data docs80.8%
30gemma2-27b· Stanford HELM specialized suites data docs80.8%
31DBRX-instruct· Stanford HELM specialized suites data docs80.4%
32Phi-3 Small 8K Instruct· Stanford HELM specialized suites data docs80.4%
33GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs80%
34GPT-4 Turbo· Stanford HELM specialized suites data docs80%
35Qwen1.5-14B· Stanford HELM specialized suites data docs80%
36DeepSeek LLM 67B Chat· Stanford HELM specialized suites data docs79.6%
37Llama 2 70b· Stanford HELM specialized suites data docs79.6%
38Claude Haiku 3.5· Stanford HELM specialized suites data docs79.2%
39GPT-4o mini· Stanford HELM specialized suites data docs78.8%
40Mistral Small· Stanford HELM specialized suites data docs78.8%
41Nova Litev1· Stanford HELM specialized suites data docs78.8%
42Claude Instant 1.2· Stanford HELM specialized suites data docs78.4%
43gemma-2-9b· Stanford HELM specialized suites data docs78%
44Snowflake Arctic Instruct· Stanford HELM specialized suites data docs78%
45Jamba 1.5 Large· Stanford HELM specialized suites data docs77.1%
46Meta Llama 3 8b· Stanford HELM specialized suites data docs77.1%
47Open Mistral Nemo· Stanford HELM specialized suites data docs77.1%
48Jamba 1.5 Mini· Stanford HELM specialized suites data docs75.5%
49Command R Plus· Stanford HELM specialized suites data docs75.1%
50GPT-3.5 Turbosetting not named by the source· Stanford HELM specialized suites data docs75.1%

Showing 1–50 of 68 results.

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