McGill-NLP/SpeechJBB
收藏资源简介:
SpeechJBB是一个用于评估大型音频语言模型(LALMs)在多种语言和语码转换语音下的安全对齐和理解能力的音频基准。它旨在测试模型在单语语音、语码转换语音以及带有自然伪词混淆的语码转换语音下是否始终拒绝有害的语音请求。该基准将JailbreakBench扩展到多种语言语音中,包含英语、德语、西班牙语、法语和意大利语的有害和良性提示的翻译和语码转换版本,并包括在安全关键术语附近插入音系合理的伪词的不同插入比例的增强变体。此外,它还包含相同语言的MGSM、Fleurs和Fleurs-SLU SIB音频文件。该数据集用于受控的安全评估、红队研究和多语言音频模型鲁棒性分析。
SpeechJBB is an audio benchmark for evaluating the safety alignment and comprehension capabilities of large audio language models (LALMs) across multiple languages and code-switched speech. It aims to test whether models consistently reject harmful speech requests in monolingual speech, code-switched speech, and code-switched speech with natural pseudo-word confusions. This benchmark extends JailbreakBench to multilingual audio scenarios, containing translated and code-switched versions of harmful and benign prompts in English, German, Spanish, French, and Italian, as well as augmented variants with varying insertion rates of phonologically plausible pseudo-words near safety-critical terms. Additionally, it includes MGSM, Fleurs, and Fleurs-SLU SIB audio files in the same languages. This dataset is used for controlled safety evaluation, red team research, and robustness analysis of multilingual audio models.




