anonymous-p2i/Planning2Interaction
收藏资源简介:
Planning2Interaction (P2I) 是一个面向恢复的多轮函数调用数据集和基准。与大多数关注成功轨迹的工具使用数据集不同,P2I专为代理在执行偏离预期路径后必须恢复的场景设计。这些偏离可能包括显式工具错误、无效状态、过时标识符或仅在后续轮次中显现的遗漏先决条件操作。P2I包含两个资源:P2I-67k(一个包含67,086条多轮函数调用轨迹的训练语料库)和P2I-Eval(一个用于直接评估函数调用恢复的保留基准)。数据集通过计划优先的构建协议创建,每条轨迹都基于一个结构化参考计划,该计划指定了工具依赖、状态流、预期结果和交互合成前的恢复目标。P2I旨在支持多轮函数调用、工具使用代理、面向恢复的监督、状态跟踪和依赖管理、工具错误或不完整执行下的代理鲁棒性以及恢复行为的诊断评估等研究。
Planning2Interaction (P2I) is a recovery-oriented dataset and benchmark for stateful multi-turn function calling. Unlike most tool-use datasets that focus on clean successful trajectories, P2I is designed for settings where an agent must recover after execution has deviated from the intended path. These deviations may include explicit tool errors, invalid state, stale identifiers, or omitted prerequisite actions that only become visible in later turns. P2I contains two resources: P2I-67k (a training corpus of 67,086 multi-turn function-calling trajectories) and P2I-Eval (a held-out benchmark for directly evaluating function-call recovery). The dataset is built with a plan-first construction protocol, where each trajectory is grounded in a structured reference plan that specifies tool dependencies, state flow, expected outcomes, and recovery targets before interaction synthesis. P2I is intended for research on multi-turn function calling, tool-use agents, recovery-oriented supervision, state tracking and dependency management, agent robustness under tool errors or incomplete execution, and diagnostic evaluation of recovery behavior.





