Jack-ppkdczgx/speechllm-gaslighting-benchmark

该数据集名为Speech LLM Gaslighting Negation Benchmark，旨在研究针对语音语言模型的煤气灯式否定提示。它包含五个语音任务：MELD（情感分类）、MMAU（音频推理）、MMSU（口语多选问答）、OpenBookQA（口语多选问答）和VocalSound（声音分类）。所有音频均标准化为16 kHz，每个示例包含一个干净提示、五个煤气灯提示变体、真实标签以及嵌入HF导出的自包含音频。导出的数据集已按源数据集模板对齐：MMSU和OpenBookQA使用口语多选问答提示，而MELD、MMAU和VocalSound使用闭集分类/选择提示。发布的导出中，每个样本包含所有五个煤气灯提示变体，ground_truth/ground_truth_text与样本的答案空间匹配，音频以16 kHz存储。保留的五种煤气灯提示类型包括：愤怒、讽刺、认知、隐性和专业。数据集总示例数为10,730，分布在五个任务中。

The dataset, named Speech LLM Gaslighting Negation Benchmark, is designed for studying gaslighting-style negation prompts against speech-language models. It includes five speech tasks: MELD (emotion classification), MMAU (audio reasoning), MMSU (spoken multiple-choice QA), OpenBookQA (spoken multiple-choice QA), and VocalSound (vocal sound classification). All audio is normalized to 16 kHz, and each example contains one clean prompt, five gaslighting prompt variants, the ground-truth label, and self-contained audio in the embedded HF export. The exported dataset has been aligned by source dataset template: MMSU and OpenBookQA use spoken multiple-choice QA prompts, while MELD, MMAU, and VocalSound use closed-set classification/selection prompts. In the released export, every sample contains all five gaslighting prompt variants, ground_truth/ground_truth_text matches the answer space for the sample, and audio is stored at 16 kHz. The five retained gaslighting prompt types are: anger, sarcasm, cognitive, implicit, and professional. The total number of examples is 10,730, distributed across the five tasks.