jAEhEEkIM/operationbench-credit-card-fraud-abrupt

该数据集是一个经过处理的OperationBench玩具流，用于一个代理概念漂移任务。环境在每个时间点发出强制周期性观察包；代理在干预成本下决定如何响应。数据集基于credit_card_fraud源数据集，任务ID为credit-card-fraud-abrupt，角色是用于突然和标签偏移的压力测试。漂移配置文件为abrupt_and_label_shift，总行数为284807，观察批次大小为2048，目标指标为F1，目标阈值为0.5。数据模式遵循共享的OperationBench玩具流模式，包括特征列从feature_00到feature_29，未使用的特征列为空。数据集源自Hugging Face的David-Egea/Creditcard-fraud-detection数据集，在固定修订版90cdcb82de30f04d084599dc37fa054776862e8e下处理。重要限制是存在极端类别不平衡，因此准确率无意义，建议使用F1、召回率、精确度或PR-AUC类指标。This dataset is a processed OperationBench toy stream for one agentic concept-drift task. The environment emits a forced periodic observation packet at every tick; the agent decides how to respond under intervention cost. It is based on the source dataset credit_card_fraud with task id credit-card-fraud-abrupt, serving as an abrupt and label-shift stress test. The drift profile is abrupt_and_label_shift, with 284807 rows, an observation batch size of 2048, target metric of F1, and target threshold of 0.5. The schema uses the shared OperationBench toy stream schema, including feature columns from feature_00 to feature_29, where unused feature columns are null. The dataset is processed from the Hugging Face David-Egea/Creditcard-fraud-detection dataset at pinned revision 90cdcb82de30f04d084599dc37fa054776862e8e. An important limitation is extreme class imbalance, making accuracy meaningless; thus, F1, recall, precision, or PR-AUC-style targets are recommended.