WenHang/GAIS
收藏资源简介:
GAIS(Grounded Agentic Interaction Synthesis)数据集旨在扩展代理智能能力,通过提供高保真、长周期的用户-助手交互轨迹,支持代理与多样化现实世界工具交互以完成复杂任务。数据集包含两个主要子集:gais_think.json(包含显式内部推理轨迹的交互)和gais_nonthink.json(无显式推理步骤的直接动作生成交互)。每个实例模拟一个部分可观察马尔可夫决策过程(POMDP),包括工具定义(从MCP服务器实现转换的可执行Python工具)、多轮对话(系统策略、用户意图、代理响应和观察结果)。数据生成通过自动化流程实现,包括MCP服务器转换、复杂依赖规划和对抗场景注入,以及用户模拟代理与助手代理的多轮对话模拟。
The GAIS (Grounded Agentic Interaction Synthesis) dataset provides high-fidelity, long-horizon user-assistant interaction trajectories designed to scale agentic capabilities, supporting general agentic intelligence in interacting with diverse real-world tools to complete complex tasks. The dataset is divided into two subsets: gais_think.json (interaction trajectories with explicit internal reasoning traces) and gais_nonthink.json (direct action generation trajectories without explicit reasoning steps). Each instance represents a simulated multi-turn Partially Observable Markov Decision Process (POMDP), including tools (executable Python tools transformed from MCP server implementations), conversations (system policy, human intent, agent response, and observation). The data is generated via an automated pipeline involving MCP server transformation, complex-dependency planning with adversarial injection, and agentic interaction simulation.




