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 Detection research benchmarks – Page 2 · SOTA2 Research

Safety Detection

Benchmarks

Dataset NameSOTA MethodSort

Dataset NameSOTA methodMetricTrendResultsLast Updated

ToxicChat (held-out)

87.7AUROC

5

May 21, 2026

BeaverTails (held-out)

MultiLayer-Linear

76.7AUROC

5

May 21, 2026

64.3TPR @ 3% FPR

5

May 21, 2026

Human-labeled evaluation set

12B Curriculum Judge

92.86Unsafe Recall (HarmBench)

4

Jun 9, 2026

TS-Bench AgentDojo-Traj (eval)

1.36Efficiency (s/sample)

4

Feb 26, 2026

TS-Bench AgentHarm-Traj (eval)

1.36Latency (s/sample)

4

Feb 26, 2026

XSTest benign safety-adjacent

Final-token probe

23False Positive Rate

3

May 14, 2026

SorryBench wrapped with HarmBench templates (held-out)

Final-token probe

86.9Detection Rate

3

May 14, 2026

SorryBench clean condition (held-out)

Final-token probe

95Detection Rate

3

May 14, 2026

AegisSafetyTest

Granite Guardian 3.3

87F1 Score

3

Mar 10, 2026

Granite Guardian 3.3

65F1 Score

3

Mar 10, 2026

Safety Evaluation Strict Safety Mode

Adaptive Abstention System

50Precision

1

Feb 26, 2026

Page 2 of 2

Popular datasets

MAS Polyguard XSTest ToxicChat Sorrybench

Follow for updates

Join SOTA2 Contact