Training dataset of the selected ALMA Cycle 9 bandpass calibration solutions

This dataset contains the bandpass calibration data and associated labels used to train and evaluate the machine-learning classifier for ALMA bandpass calibration anomalies described in Xue et al. (in preparation). The dataset is constructed from selected public ALMA Principal Investigator (PI) science observations delivered during Cycle 9 (October 2022 - September 2023). It contains 84,572 bandpass calibration solutions from 303 execution blocks (EBs), belonging to 151 member observing unit sets (MOUSs). The individual solutions are aggregated at the polarization-pair level, with anomaly labels combined using a logical OR operator across the two polarizations. The resulting dataset contains 42,286 polarization-pair samples. The dataset is provided in CSV format containing 42,286 rows and 27 columns. Each row corresponds to one polarization pair for a given EB, antenna and spectral window. The columns include: Data identifiers: eb_uid, spw_name_ms, and antenna_name, identifying the EB, spectral window, and antenna associated with each sample. Classification information: ground_truth contains the Boolean reference label used for classifier training and evaluation, and xgboost_prediction contains the corresponding prediction from the XGBoost classifier. True represents polarization pairs containing bandpass amplitude anomalies and False represents nominal, non-anomalous solutions. Bandpass calibration solutions: amplitude_0 and amplitude_1 contain the bandpass amplitude arrays for the two polarizations, with the corresponding frequency arrays stored in frequency_array_0 and frequency_array_1. Data flags and atmospheric information: flag_array_0 and flag_array_1 contain the channel-level flags associated with each solution. atmospheric_interference_0 and atmospheric_interference_1 identify channels associated with modeled atmospheric interference. Features: eight features used as inputs to the classifier: score_masked_0/1: scan-statistic score ranging from 0 to 1, with larger values indicating a spectral interval deviating more from the baseline, derived with spectral edges and atmospheric interference masked; dimensionless. segment_width_masked_0/1: frequency range of the spectral interval with the highest masked scan-statistic score; units of Hz. score_unmasked_0/1: scan-statistic score ranging from 0 to 1, with larger values indicating a spectral interval deviating more from the baseline, derived without masking atmospheric interference; dimensionless. segment_width_unmasked_0/1: frequency range of the spectral interval with the highest unmasked scan-statistic score; units of Hz. score_fixed_0/1: scan-statistic score ranging from 0 to 1, with larger values indicating a spectral interval deviating more from the baseline, derived with a fixed interval width of 62.5 MHz and without masking atmospheric interference; dimensionless. amp_norm_nmad_diff4_0/1: normalized dispersion of the bandpass amplitude calculated with a four-channel separation; dimensionless. channel_spacing: spectral interval between adjacent channels, computed from the frequency array; units of Hz. receiver_band: ALMA observational frequency band; categorical feature. Subscripts (0, 1) correspond to features derived separately from the two orthogonal polarization solutions; features common to both polarizations have no subscript. If you are interested in using this dataset, please contact Ci Xue or a member of the CosmicAI Observable Working Group. Please acknowledge use of this dataset by citing both this Zenodo record and the article: C. Xue, B. S. Mason, G. A. Rakib, T. Ashton, J. M. Phillips, R. A. Loomis, I. Yoon, J. Hibbard, I. Toledo, and E. J. Murphy, Classifier Framework for ALMA Bandpass Calibration Anomalies (in preparation). (We will update the reference when the paper is published.) (08/16/2026)