SorryBench
Adversarial prompts with linguistic mutations for testing safety.
About SorryBench
SORRY-Bench is a dataset of unsafe instructions intended for evaluating large language model safety refusal behavior. The dataset is published by its authors under the sorry-bench name on Hugging Face, with Tinghao Xie listed as the dataset card contact. LLM developers can use the dataset to check model refusal behaviors on each of its 44 safety categories, or on a subset of them according to their own safety policies.
The dataset covers risky topics across four high-level domains: Hate Speech Generation, Assistance with Crimes or Torts, Potentially Inappropriate Topics, and Potentially Unqualified Advice. Jailbreak researchers can use it to benchmark the effectiveness of different jailbreaking attacks and defenses.
Drafted from huggingface.co · not yet confirmed by SorryBench
How it works
SorryBench tests how a model refuses across 44 fine grained unsafe topics, with 440 base prompts and then 20 linguistic mutations of each: persuasion, technical terms, role play, misspellings, translations into other languages, and so on. That gives you thousands of variants and shows whether refusals hold when the request is phrased differently. It ships with a fine tuned judge for scoring.
It's the benchmark to use when the question is "does phrasing break our guardrail."
From the project
Reviews
Be the first to review it
AnonymousKnow a team weighing SorryBench? Send them this page.