UVCGAN-S: Pre-processed sPHENIX Dataset

sPHENIX Calorimeter Data for UVCGAN-S Training Overview This dataset provides preprocessed HDF5 files derived from sPHENIX detector simulations for training and evaluating the UVCGAN-S background subtraction model described in: "Robust and Generalizable Background Subtraction on Images of Calorimeter Jets using Unsupervised Generative Learning" arXiv: [https://arxiv.org/abs/2510.23717] The data consists of calorimeter energy histograms projected onto eta-phi space, converted from ROOT TH2D histograms to HDF5 format for efficient machine learning workflows. Source Data This dataset was derived from the ROOT files available at: https://zenodo.org/records/17594612 The original ROOT dataset contains detailed documentation of simulation parameters, event generators (PYTHIA, HIJING, JEWEL), and physics context. Dataset Structure ```2025-06-05_jet_bkg_sub/├── train/│ ├── signal.h5 -> type11_run19_jet30_pythia_noNoise_allCentrality.h5│ ├── background.h5 -> type4_run19_hijingAll_noNoise_cent0.h5│ ├── embed.h5 -> type11plus4_run19_jet30_hijing_noNoise_cent0.h5│ └── *.h5_index.csv (index files for each HDF5)├── val/│ ├── embed.h5 -> type11plus4_run19_jet30_hijing_noNoise_cent0.h5│ └── *.h5_index.csv└── test/ ├── embed.h5 -> type4_hijing_plus_jewel_jet30_40_50_noNoise_cent0.h5 └── *.h5_index.csv``` Symbolic links provide semantic naming while preserving detailed filenames that encode generation parameters. Data Splits Training Set - **Signal** (`signal.h5`): 2,635,582 isolated PYTHIA jets (pT > 30 GeV)- **Background** (`background.h5`): 985,791 HIJING events (0-10% centrality Au+Au)- **Embedded** (`embed.h5`): 633,000 PYTHIA jets embedded in HIJING backgrounds Validation Set - **Embedded** (`embed.h5`): 200,000 events (randomly split from same generation as training embedded data) Test Set - **Embedded** (`embed.h5`): 833,000 JEWEL jets embedded in HIJING backgroundsCritical: JEWEL jets are out-of-distribution (quenched jets with jet-medium interactions). Tests model generalization to unseen event generator physics HDF5 File Format `data` Dataset - **Shape**: `(N, 24, 64)` where N is the number of events- **Type**: Float32- **Content**: Calorimeter energy histograms in eta-phi space - Dimension 0 (24 bins): Pseudo-rapidity (eta) - Dimension 1 (64 bins): Azimuthal angle (phi) - Values represent summed electromagnetic + hadronic calorimeter energies Index CSV Files Each HDF5 file has an accompanying `*.h5_index.csv` file with columns:- `sample`: HDF5 row index- `file`: Source ROOT filename- `hist`: ROOT histogram name (format: `h_eta_phi_cent0_fileXX_evtYY;1`) These enable traceability back to original ROOT histograms and event matching for paired evaluation.