FlowBench
收藏资源简介:
FlowBench是由浙江大学和阿里巴巴集团合作创建的,旨在通过系统化的基准测试评估基于LLM的代理在任务规划中的表现。该数据集包含51个不同场景,覆盖6个领域,采用多种知识格式(文本、代码、流程图)以适应实际应用。创建过程分为任务收集、工作流组织和会话生成三个阶段,确保数据集的多样性和专业级标注。FlowBench适用于评估代理在复杂任务中的规划能力,特别是在需要外部工作流知识支持的场景中,为未来代理规划研究提供了挑战性的基准。
FlowBench is a collaborative benchmark developed by Zhejiang University and Alibaba Group, designed to evaluate the task planning performance of LLM-based AI agents through systematic benchmark testing. This dataset comprises 51 distinct scenarios spanning 6 domains, and supports multiple knowledge formats including text, code, and flowcharts to cater to real-world applications. Its development process is divided into three stages: task collection, workflow organization, and session generation, which guarantees the dataset's diversity and professional-grade annotations. FlowBench is tailored to evaluate the planning capabilities of AI agents in complex tasks, particularly in scenarios requiring external workflow knowledge support, providing a challenging benchmark for future research in agent planning.

- 1FlowBench: Revisiting and Benchmarking Workflow-Guided Planning for LLM-based Agents浙江大学, 阿里巴巴集团 · 2024年



