AI45Research/AgentDoG1.0-Training-Data

AgentDoG1.0训练数据集是一个用于多步骤、工具使用AI代理安全建模的监督指令调优数据集。它包含两个主要配置：AgentDoG-BinarySafety用于轨迹级安全分类（判断代理轨迹是否安全，输出safe或unsafe），包含4000条平衡的训练记录（2000条安全，2000条不安全）；AgentDoG-FineGrainedTaxonomy用于细粒度不安全轨迹诊断，基于ATBench分类法输出风险来源、失败模式和现实世界危害三个维度的标注，包含4000条训练记录和1044个独特的分类输出组合。数据集采用指令调优格式，每个样本包含instruction（完整任务提示和代理轨迹）、input（当前版本为空）和output（目标标签或分类标注）。该数据集旨在用于监督微调、消融研究、标签空间分析和诊断模型比较，与ATBench基准测试配套使用。AgentDoG1.0 Training Data is a supervised instruction-tuning dataset for trajectory-level safety modeling of multi-step, tool-using AI agents. It includes two configurations: AgentDoG-BinarySafety for trajectory-level safety classification (judging whether an agent trajectory is safe or unsafe, outputting safe or unsafe), containing 4,000 balanced training examples (2,000 safe and 2,000 unsafe); and AgentDoG-FineGrainedTaxonomy for fine-grained diagnosis of unsafe trajectories, outputting annotations across three dimensions of the shared ATBench taxonomy: Risk Source, Failure Mode, and Real-World Harm, containing 4,000 training examples with 1,044 unique taxonomy-output combinations. The dataset follows an instruction-tuning schema with each sample containing an instruction (full task prompt and agent trajectory), input (empty in the current release), and output (target label or taxonomy annotation). It is intended for supervised fine-tuning, ablation studies, label-space analysis, and diagnostic model comparison, and is paired with the ATBench benchmark for evaluation.