Hunting for vampires and other unlikely forms of parity violation at the Large Hadron Collider: truth-jet and reco-jet datasets

All truth-jet and reco-jet datasets used in the paper.

Main: fragments are named `truth-jet_${MODEL}.tar.gz`, where MODEL is either `pv_msme_${lambdaPV}` (PV-mSME and lambdaPV is a floating point number with "." replaced with "p"), or `sm` for the Standard Model (lambdaPV = 0).

Within each are three directories for the independent splits:

• train,
• test, and
• private_test.

 In the paper, we describe "test" as the validation set and "private_test" as the test set. Data in "private_test" were not used until models were finalized for the paper.
Within each of those are the processed results from simulations with different random seeds. They are independent shards which can be trivially combined.

Each data file is in h5 format. Its data are under the key "events" as an array with shape (n, 20). That last axis contains the reconstructed four-momenta of the hardest five jets in the order [Px, Py, Pz, E] * 5, with missing jets filled with zeros.

Truth-jet files have additional keys "flavors" and "helicities". These contain truth-level flavour and helicity information, respectively. Their shape is (n, 5), where the last axis contains the zero-padded results. Flavours are encoded in the PDG ID scheme.

Rotated: Rotated PV-mSME data are in archives named `truth-jet-rot_${HOUR}.tar.gz`.
These have similar (train, test, private_test) structures as the others, but comprise parts of the same mixed model and should be combined by subsampling to a weighted average.

HOUR is an integer rotation in [0, 23] for which the detector is rotated by an angle given in radians as HOUR * 2 pi / 24. This rotated dataset has lambdaPV=1. We demonstrate merging of these datasets in the code sharing (git, Zenodo).