ari-abb/ssh-conversation-risk-dataset

SSH对话风险数据集是一个合成的多轮对话数据集，模拟用户与AI助手之间的对话，用于在对话级别和轮次级别标注自杀和自残（SSH）风险。数据集包含97个对话，涵盖7种风险模式，旨在训练和评估能够捕捉逐渐升级模式的对话级SSH风险分类器。数据集通过两代理LLM管道生成，包括逐步升级、模糊隐藏信号、假设框架等风险模式，并提供了严重程度量表（0-5）和轨迹类型。数据集适用于训练现代BERT、Longformer等模型，用于评估护栏系统在逐步升级和假设框架上的表现。

The SSH Conversation Risk Dataset is a synthetic multi-turn conversation dataset between simulated users and an AI assistant, annotated for suicide and self-harm (SSH) risk at both the turn level and conversation level. It consists of 97 conversations across 7 risk profiles, designed to train and evaluate conversation-level SSH risk classifiers that detect gradual escalation patterns. Generated via a two-agent LLM pipeline, the dataset includes risk profiles like Gradual Escalation, Ambiguous Buried Signal, and Hypothetical Framing, with a severity scale (0–5) and trajectory types. It is intended for training models like ModernBERT and Longformer, and for evaluating guardrail systems on gradual escalation and hypothetical framing.