AI45Research/2026_summer_camp_teseset
收藏资源简介:
本仓库提供AgentDoG-Lite:基于AgentDoG 1.5的轻量级Agent安全诊断挑战的测试数据,用于评估参赛方案在Agent轨迹级安全判断任务上的表现。与传统只判断单轮输入或最终回复是否安全的评测不同,本测试集关注完整Agent执行轨迹中的安全风险。每条样本包含一段多轮Agent轨迹,可能涉及用户请求、工具描述、工具调用、工具返回、环境反馈和最终回复。参赛系统需要根据完整轨迹判断Agent是否执行了不安全行为或不安全决策模式。本测试集主要用于评测参赛者基于Qwen3.5-0.8B构建的安全判断系统,适用于单人赛的training-free harness方法和团队赛的基于Qwen3.5-0.8B微调后的模型。核心输出为二分类安全判断,包括safe或unsafe。本数据集不应用于训练或微调模型、few-shot示例选择、prompt搜索或测试集调参、生成真实攻击流程或执行危险命令,以及作为真实部署环境中的唯一安全判断依据。
This repository provides AgentDoG-Lite: test data for the lightweight Agent security diagnosis challenge based on AgentDoG 1.5, which is used to evaluate the performance of participating solutions on the Agent trajectory-level security judgment task. Unlike traditional evaluations that only judge whether single-round inputs or final responses are safe, this test set focuses on security risks in the complete execution trajectory of Agents. Each sample contains a multi-round Agent trajectory, which may involve user requests, tool descriptions, tool calls, tool returns, environmental feedback, and final responses. Participating systems are required to judge whether the Agent has executed unsafe behaviors or unsafe decision-making patterns based on the complete trajectory. This test set is mainly used to evaluate the security judgment systems built by participants based on Qwen3.5-0.8B, and is applicable to the training-free harness method for individual competitions and the fine-tuned models based on Qwen3.5-0.8B for team competitions. The core output is binary classification security judgment, including safe or unsafe. This dataset must not be used for model training or fine-tuning, few-shot example selection, prompt search or test set tuning, generating real attack processes or executing dangerous commands, or serving as the sole security judgment basis in real deployment environments.




