JailbreakBench
Harmful behaviors used as a benchmark for jailbreaking evaluation.
About JailbreakBench
JailbreakBench publishes the JBB-Behaviors dataset on Hugging Face. The dataset contains a behaviors subset of 200 rows and a judge_comparison subset of 300 rows.
The harmful split has 100 rows and the benign split has 100 rows. The dataset page cites arXiv paper 2404.01318.
Drafted from huggingface.co · not yet confirmed by JailbreakBench
How it works
JailbreakBench is a standardized benchmark for jailbreak attacks and defenses. It provides 100 harmful behaviors and 100 benign ones, each tied to a category and a source, plus a leaderboard, an attack artifact repository, and a judge protocol for scoring whether a model complied. The point is reproducibility: everyone evaluates the same behaviors the same way.
You'd use it to test a model or a guardrail against known jailbreak techniques and to compare your results to published ones.
From the project
Reviews
Be the first to review it
AnonymousKnow a team weighing JailbreakBench? Send them this page.