ServiceNow/Dr-CiK
收藏资源简介:
Dr-CiK是一个用于评估智能体能力的基准测试数据集,专注于时间序列预测任务。其核心目标是测试智能体是否能够从包含噪声的文档语料库中检索与预测相关的上下文信息,过滤掉干扰文档,将检索到的上下文提炼为可用于预测的证据,并基于这些证据生成预测。该数据集模拟了现实世界的时间序列预测场景,其中预测不仅依赖于历史观测数据,还需要从异构、嘈杂的信息源中主动发现外部上下文。与现有通常假设支持性上下文已提供的基准测试不同,Dr-CiK的每个任务都将一个时间序列与一个包含支持性文档和干扰文档的语料库配对,要求智能体自行寻找并使用正确的证据。数据集包含279个任务、10,342个文档(其中3,367个支持性文档,6,975个干扰文档),并分为公开开发集(199个合成任务)和隐藏测试集(80个人工编写任务)。每个任务涉及一个4跳推理链,干扰文档分为五种子类型:混淆型、噪声型、时间序列误导型、配置文件不匹配型和时序错误型。
Dr-CiK is a benchmark for evaluating whether agents can retrieve forecasting-relevant context from a noisy document corpus, filter out distractors, distill the retrieved context into forecast-useful evidence, and produce forecasts grounded in that evidence. It addresses real-world time-series forecasting scenarios where forecasting depends not only on historical observations but also on external context that must be actively discovered from heterogeneous, noisy information sources. Unlike existing context-aided forecasting benchmarks that typically assume supporting context is already provided, Dr-CiK removes that assumption: each task pairs a time series with a corpus of supporting and distractor documents, and the agent must find and use the right evidence on its own. The dataset includes 279 tasks, 10,342 documents (3,367 supporting, 6,975 distractor), split into a public dev set (199 synthetic tasks) and a hidden test set (80 human-authored tasks). Each task is built around a 4-hop reasoning chain, with distractor documents categorized into five subtypes: confounder, noisy, timeseries, profile, and temporal.




