Scheduling the Solver: Data for Learning-Guided Structure-Aware Solver Control in Job Shop Scheduling

This dataset contains the research data and minimal reproducibility artifacts supporting the manuscript “Scheduling the Solver: Learning-Guided Structure-Aware Solver Control for Job Shop Scheduling”. The deposit includes: (i) the 54-family development data used for the strict out-of-fold H4B learning stage, including structural descriptors, realized action gaps, advantage targets, fold assignments, and out-of-fold predictions; (ii) a compact summary of the independent H4C confirmation on previously unseen families; (iii) the frozen protected JSPLIB benchmark manifest; (iv) consolidated per-instance external evaluation results for 10, 60, and 600 second budgets; (v) the held-out permutation-sensitivity data used in the feature-importance analysis; and (vi) a minimal Python script for reproducing the main H4B and external-evaluation summaries. The protected JSPLIB benchmark was not used for model training or model selection. The final structure-aware policy was trained only on the development set, while the independent H4C confirmation set was kept separate from training. The external evaluation compares the proposed structure-aware CP-SAT control strategy with fixed CP-SAT configurations, IBM CP Optimizer, and Gurobi. This first public release intentionally contains the data required to support the main empirical claims of the manuscript, rather than the complete experimental development archive or raw solver logs. Additional material may be released in subsequent versions if required for further reproducibility or review.