AliDjl/Multi-Turn-Insurance-Underwriting

该数据集包含商业保险核保人与AI助手之间多轮交互的样本轨迹及相关元数据。系统采用langgraph框架、模型上下文协议以及ReAct智能体构建。每个样本中，核保人需完成与小型企业近期保险申请相关的具体任务。数据集涵盖了6种不同类型的任务，部分任务包含涉及更细致、复杂核保逻辑的子任务。这些任务平均需要3-7步推理和工具调用，对话总轮数在10-20轮之间。数据集由Snorkel AI策划，采用Apache-2.0许可证，旨在支持对大型语言模型在多轮任务（需要正确推理、工具调用和与用户交互）上的评估。场景设置具有可验证的答案，同时足够复杂，对最先进的LLM具有挑战性。This dataset contains sample trajectories and associated metadata for multi-turn interactions between commercial insurance underwriters and AI assistants. The system is built using the LangGraph framework, model context protocol, and ReAct AI Agent. In each sample, the underwriter is tasked with completing specific tasks related to recent insurance applications from small businesses. The dataset includes six distinct task types, with some tasks containing subtasks that involve more granular and complex underwriting logic. These tasks typically require 3 to 7 steps of reasoning and tool invocation, with the total number of dialogue turns ranging from 10 to 20. Curated by Snorkel AI and released under the Apache-2.0 license, this dataset is designed to support the evaluation of large language models (LLMs) on multi-turn tasks that demand accurate reasoning, tool invocation, and user interaction. The scenarios feature verifiable answers while being sufficiently complex to pose challenges to state-of-the-art large language models.