Categories Companies Research

Loading the SOTA2 catalog…

Explore

Categories Companies Research

Contribute

Join SOTA2 Contact

SOTA2

© 2026 SOTA2

Build in Bellevue, USA

WizWand is now SOTA2 Research

Papers Benchmarks Datasets Tasks

Safety Evaluation research benchmarks – Page 11 · SOTA2 Research

Safety Evaluation

Benchmarks

Dataset NameSOTA MethodSort

Dataset NameSOTA methodMetricTrendResultsLast Updated

Safety Evaluation

Llama-3.3-70B-Instruct

73.97Overall Safety Score

9

Mar 18, 2026

0.8Over-Refusal Rate

9

Mar 13, 2026

MultiJail 10 langs

Tiny Aya Global

87Min Safe Rate

9

Mar 13, 2026

Prefilling Attacks Qwen2.5-14B-Instruct

11.5ASR

9

Feb 26, 2026

MultiJail-AR (test)

89.101Safety Rate (%)

9

Feb 26, 2026

MultiJail-KO (test)

90.053Safety Rate

9

Feb 26, 2026

MultiJail EN (test)

93.651Safety Rate

9

Feb 26, 2026

Do-Not-Answer (test)

3.195ASR

9

Feb 26, 2026

SafetyBench (test)

81.321Accuracy

9

Feb 26, 2026

99.3Safety Score

9

Feb 26, 2026

93.2Safety Score

9

Feb 26, 2026

98.3Safety Score

9

Feb 26, 2026

WildChat unsafe prompts

99.82Not-Unsafe Rate

9

Feb 26, 2026

Norwegian pack (full 36-scenario)

43.68Judge-L Score

8

May 8, 2026

LM Safety Evaluation Dataset

0Unsafe Rate

8

Apr 20, 2026

HarmBench first 50 behaviors (fixed subset)

Re-mask only (ablation)

0ASR

8

Apr 13, 2026

99.4Safety Score

8

May 28, 2026

Claude 3.5 Sonnet

4.5ASR

8

Feb 26, 2026

Claude 3.5 Sonnet

12ASR

8

Feb 26, 2026

3.5ASR

8

Feb 26, 2026

0Collision Count

8

Feb 26, 2026

0Collision Rate

8

Feb 26, 2026

0Collision Rate

8

Feb 26, 2026

0Collision Count

8

Feb 26, 2026

HEx-PHI-FigStep

1Safety Score (1-ASR)

8

Feb 26, 2026

Page 11 of 17

Popular datasets

XSTest HarmBench BeaverTails HEX-PHI AdvBench

Follow for updates

Join SOTA2 Contact