Mora / Datasets / McGill-NLP/SpeechJBB

SpeechJBBFree and public

SpeechJBB is an audio benchmark for evaluating safety alignment and comprehension in large audio language models (LALMs) under multilingual and code-switched speech. It is designed to test whether models consistently refuse harmful spoken requests when prompts are expressed in monolingual speech, code-switched speech, and code-switched speech with natural-sounding pseudo-word obfuscation. The benchmark extends JailbreakBench into… See the full description on the dataset page: https://huggingface.co/datasets/McGill-NLP/SpeechJBB.