ARFBench/ARFBench

ARFBench（异常推理框架基准）是一个多模态时间序列推理基准，包含750个问答对，这些问答对基于从可观测性平台供应商收集的真实世界事件数据构建。数据集由142个唯一时间序列组成，这些时间序列来自63个不同的事件讨论线程，时间序列在变量数量上差异显著。收集的时间序列涵盖软件事件响应和可观测性中的多个领域，包括应用使用（如服务器请求）、基础设施（如CPU使用率等计算机系统测量）、网络（如延迟、可用性）、数据库（查询执行时间、查询错误率）以及安全（如漏洞警报）。数据集结构包括一个CSV文件，其中包含问题、任务类别、难度、选项、正确答案、查询组和插值标志等列；每个唯一时间序列还关联两种数据模态：时间序列数据（最多6个不同间隔）和时间序列图（由Matplotlib或Plotnine生成的PNG图像）。

ARFBench (Anomaly Reasoning Framework Benchmark) is a multimodal time-series reasoning benchmark consisting of 750 question-answer (QA) pairs composed from real-world incident data collected at an observability platform vendor. The dataset consists of 142 unique time series collected from 63 different incident discussion threads, with time series varying significantly in the number of variates. The time series span many domains in software incident response and observability, including application usage (e.g., server requests), infrastructure (e.g., measurements of computer systems such as CPU usage), networking (e.g., latency, availability), database (query execution time, query error rates), and security (e.g., vulnerability alerts). The dataset structure includes a CSV file with columns for Question, Task Category, Difficulty, Options, Correct Answer, Query Group, and Interpolation flags; each unique time series also has two associated modalities: time series data (up to 6 different intervals) and time series plots (PNG images generated by Matplotlib or Plotnine).