dovzhikova/HiddenSignals_v1

HiddenSignals-v1是一个专门设计的语料库，旨在解决当前AI安全模型中的临床差距问题。与关注明确临床术语的标准数据集不同，该数据集汇总了隐晦、俚语和逃避性的痛苦信号（如我要退出了、sewerslide、买张去瑞士的票）。数据通过匿名同伴支持平台MindBridge收集，并由临床心理学研究团队使用哥伦比亚自杀严重程度评定量表（C-SSRS）进行标注。数据集的目标是使大型语言模型（LLMs）能够检测危机场景中的假阴性案例，即识别那些处于风险中但使用算法逃避技术或亚文化俚语来掩盖其意图的用户。数据集包含匿名聊天片段、识别的特定俚语术语和已验证的临床风险标签。数据集的创建遵循严格的匿名化和伦理标准，包括K-Anonymity管道和个人身份信息（PII）的清除。

HiddenSignals-v1 is a specialized corpus designed to address the Clinical Gap in current AI safety models. While standard datasets focus on explicit clinical terminology, this dataset aggregates implicit, slang-based, and evasive distress signals (e.g., Im checking out, sewerslide, buying a ticket to Switzerland). The data is collected via MindBridge, an anonymous peer-support platform, and annotated by a team of clinical psychology researchers using the Columbia-Suicide Severity Rating Scale (C-SSRS). The research goal is to enable Large Language Models (LLMs) to detect False Negatives in crisis scenarios—identifying users who are at risk but are using algorithmic evasion techniques or sub-cultural slang to mask their intent. The dataset includes anonymized chat segments, identified slang terms, and verified clinical risk labels. The creation of the dataset follows strict anonymization and ethical standards, including a K-Anonymity pipeline and the scrubbing of Personally Identifiable Information (PII).