Аңдатпа
Quality monitoring in diagnostic-device manufacturing requires datasets that preserve temporal, hierarchical, and calibration-related structure while avoiding information leakage in predictive evaluation. This study presents the Diabetes-QMS Benchmark, a de-identified multi-level industrial dataset comprising 800 records: 70 completed glucose quality-control series, 690 HbA1c measurement-level records, and 40 HbA1c production-lot records collected between 2020 and 2026. Technical validation included statistical process control, leakage-aware classification analysis, production-lot-grouped and leave-one-year-out calibration evaluation, haematocrit-stratified bias assessment, conformal prediction, and lot-level anomaly screening. Glucose QC monitoring revealed increasing instability in the monitored quality-control indicators over time, with the most pronounced excursion occurring in 2025. Under production-lot-grouped validation, Random Forest calibration achieved performance numerically comparable with factory calibration (MAE 0.506 versus 0.541), whereas leave-one-year-out evaluation showed substantially poorer temporal generalisation of the static model. Conformal coverage similarly deteriorated under temporal shift, and calibration bias varied across haematocrit ranges. Lot-level analysis revealed pronounced temporal calibration drift and increased anomaly prevalence in 2025. The benchmark therefore provides a reusable resource for investigating manufacturing quality monitoring, temporal robustness, uncertainty quantification, anomaly screening, and leakage-aware machine learning under realistic industrial data constraints.