TAPILOT-CROSSING

TAPILOT-CROSSING数据集是由香港大学和微软亚洲研究院合作创建的，旨在评估大型语言模型（LLM）代理在交互式数据分析中的能力。该数据集包含1024个交互记录，涵盖了NORMAL、ACTION、PRIVATE和PRIVATE ACTION四个实际应用场景。数据集的构建利用了DECISION COMPANY这一多代理环境，有效减少了人工干预的需求。TAPILOT-CROSSING数据集特别关注于模拟真实世界的数据分析场景，其中用户与LLM代理通过实时数据探索进行合作，以支持决策制定。此外，数据集还引入了自适应交互反射（AIR）策略，通过从成功的历史中学习，指导LLM代理进化为有效的交互式数据分析代理。

The TAPILOT-CROSSING dataset was co-created by The University of Hong Kong and Microsoft Research Asia, aiming to evaluate the capabilities of large language model (LLM) agents in interactive data analysis. This dataset contains 1024 interaction records, covering four real-world application scenarios: NORMAL, ACTION, PRIVATE, and PRIVATE ACTION. The construction of the dataset leverages the multi-agent environment DECISION COMPANY, which effectively reduces the need for manual intervention. The TAPILOT-CROSSING dataset specifically focuses on simulating real-world data analysis scenarios, where users collaborate with LLM agents through real-time data exploration to support decision-making. Furthermore, the dataset introduces the Adaptive Interaction Reflection (AIR) strategy, which guides LLM agents to evolve into effective interactive data analysis agents by learning from successful historical interactions.