agent-diff-bench
收藏资源简介:
Agent-Diff Bench 是一个用于评估代理型大型语言模型(LLMs)在通过外部API执行真实世界任务时的性能的基准测试框架。数据集包含224个企业软件工作流任务,采用80/20的训练/测试分割。其独特之处在于引入了“状态差异合约”(state-diff contract),将过程与结果分离,任务成功的定义是环境状态的预期变化是否实现,而非模糊的跟踪或参数匹配。数据集涵盖的服务包括Slack、Linear、Box和Google Calendar。评估基于状态差异,比较沙盒环境的“之前”和“之后”快照。数据集特征包括问题、答案、测试ID、测试名称、服务、任务范围、操作类型、实体范围、信息可用性、提示歧义和信息等。适用于文本生成任务,特别是涉及API交互和代理行为的场景。
Agent-Diff Bench is a benchmark framework for evaluating the performance of agentic Large Language Models (LLMs) when executing real-world tasks via external APIs. The dataset comprises 224 enterprise software workflow tasks with an 80/20 train/test split. Its distinctive feature is the introduction of the "state-diff contract", which decouples process from outcome. Task success is defined by whether the expected changes to the environment state are achieved, rather than relying on vague tracking or parameter matching. The dataset covers services such as Slack, Linear, Box, and Google Calendar. Evaluations are conducted based on state differences, comparing the "before" and "after" snapshots of the sandbox environment. The dataset attributes include questions, answers, test IDs, test names, target services, task scope, operation type, entity scope, information availability, prompt ambiguity, and associated information. It is suitable for text generation tasks, particularly scenarios involving API interactions and agent behavior.



