GrandgemMa-Corpus
收藏资源简介:
GrandgemMa-Corpus(又名ScamBench)是一个采用CC-BY 4.0许可证的多模态诈骗分类语料库,专为训练和评估面向老年人安全的设备端诈骗检测模型而构建。该数据集支持纯文本和多模态(文本+音频)两种微调模式,包含6019条样本,其中训练集5417条,留出测试集602条。类别分布为诈骗类3814条和非诈骗类2205条。从模态上看,包含4400条“文本+音频”样本和1619条纯文本样本。数据集整合了三个来源:BothBosu/multi-agent-scam-conversation(1600条,Apache-2.0许可证)、Claude生成的老年人诈骗场景(19条,CC-BY-4.0许可证)以及TeleAntiFraud多模态数据集(4400条,Apache-2.0许可证)。TeleAntiFraud来源的样本嵌入了音频(mp3格式),使模型能够学习从文本转录中丢失的韵律线索(如紧迫感、压力、语调)。数据集的列模式包括文本内容、音频数据、模态类型、类别标签、来源标识、来源许可证、来源URL、原始ID以及PII处理标记。该数据集是GrandgemMa黑客松(Gemma 4 Good,2026-05-17)项目的一部分,旨在促进针对老年人安全保护的诈骗检测模型的开发与评估。
GrandgemMa-Corpus (also known as ScamBench) is a multimodal scam classification corpus under the CC-BY 4.0 license, specifically built for training and evaluating on-device scam detection models for elderly safety. The dataset supports both text-only and multimodal (text + audio) fine-tuning modes. It contains 6019 samples, with 5417 in the training set and 602 in the held-out test set. The class distribution includes 3814 scam samples and 2205 not_scam samples. In terms of modality, it comprises 4400 text + audio samples and 1619 text-only samples. The dataset integrates three sources: BothBosu/multi-agent-scam-conversation (1600 samples, Apache-2.0 license), Claude-generated elderly scam scenarios (19 samples, CC-BY-4.0 license), and the TeleAntiFraud multimodal dataset (4400 samples, Apache-2.0 license). Samples from the TeleAntiFraud source include embedded audio (in mp3 format), enabling models to learn prosodic cues (such as urgency, stress, and intonation) that are lost in text transcription. The column schema includes text content (text, with personally identifiable information PII filtered), audio data (audio, null for some samples), modality type, category label (0 for not_scam, 1 for scam), source identifiers (source_id, source_name), source license, source URL, original ID, and PII handling markers (pii_redacted, pii_filtered). This dataset is part of the GrandgemMa hackathon (Gemma 4 Good, 2026-05-17) project, aimed at advancing the development and evaluation of scam detection models for elderly safety protection.




