TE-GvA: A Counterfactual Gather-or-Act Decision Dataset for Fault-Driven Incident Response on the Tennessee Eastman Process

TE-GvA is a dataset of gather-or-act decisions in fault-driven incident response on the revised Tennessee Eastman process. Each record is a decision point of a simulated incident, with the cost of every candidate action estimated by branched rollouts of the full simulator under common random numbers. For the ambiguity between a failed sensor link and a frozen I/O module, the dataset also provides belief labels computed from twin episodes. It contains two sets, natural and stress, with 9,908 decision points and 143,447 labelled candidate actions, split into training, validation and test. DATASHEET.md describes the contents, every field and the limitations. The code that generates, labels and evaluates the data is available at https://github.com/ruzbahani/TE-GvA.