TracerTraj
收藏资源简介:
TracerTraj数据集是AgenTracer项目的一部分,该数据集包含超过2000条高保真度的多智能体系统失败轨迹。这些轨迹是通过反事实回放和程序化故障注入自动生成的,旨在帮助理解和识别导致系统失败的关键错误步骤。该数据集用于训练AgenTracer-8B,这是一个轻量级的故障追踪器,能够在多智能体交互中高效地诊断错误。AgenTracer-8B在Who&When基准测试中表现出色,超过了其他大型LLM,为多智能体系统的自我纠正和自我进化提供了强大的支持。
The TracerTraj dataset is part of the AgenTracer project. This dataset contains over 2000 high-fidelity failure trajectories of multi-agent systems. These trajectories are automatically generated via counterfactual replay and procedural fault injection, aiming to help understand and identify the critical erroneous steps that lead to system failures. This dataset is used to train AgenTracer-8B, a lightweight fault tracker that can efficiently diagnose errors during multi-agent interactions. AgenTracer-8B performs excellently in the Who&When benchmark, outperforming other large-scale LLMs, and provides strong support for the self-correction and self-evolution of multi-agent systems.




