TopTagXL: A top quark tagging dataset for large scale studies in jet tagging

A large set of Monte Carlo simulated LHC events for the development of deep learning tools in high-energy physics. The dataset contains 135M events divided into 100M events for training, 25M events for testing, and 10M events for validation. We evenly simulate QCD jets, originating from light hadrons, and jets originating from top quarks. The training files 1 to 5 include QCD jets, while the files 6 to 10 are top jets. Events are generated with Pythia8.310, and the detector simulation is done with Delphes3.5.0 using the CMS configuration. The cards used for the generation of the scattering process and the detector simulation are included in this repository: "qcd_jets_pythia_TopTagXL.cmnd" and "top_jets_pythia_TopTagXL.cmnd" are the Pythia cards used for the generation of the scattering events. They contain a parton-level cut at 400 GeV, and top quarks are only allowed to decay hadronically; "delphes_card_CMS_TopTagXL.tcl" is the default CMS card in Delphes3.5.0 to which we added the smearing of the charged tracks impact parameter as parametrized in "trackResolutionCMS.tcl". The parametrization is already available in Delphes3.5.0 and included here for completeness. We only include fully reconstructed jets, i.e. we require $\Delta R < 0.8$ between the clustered jet kinematics and the parton-level kinematics of the top decays. We further require all jets to have $500 < p_T < 1000$ GeV and $|\eta| < 2$. Jets are clustered using FastJet with the anti-$k_T$ algorithm and a radius of $R=0.8$. The dataset contains up to 128 constituents with the following information: "part_px", "part_py", "part_pz", "part_energy": four-momenta of the constituent; "part_charge": an integer in (-1, 0, 1) indicating the charge of the constituent; a one-hot encoded particle identification number indicated with the labels "part_"+"isChargedHadron", "isNeutralHadron", "isPhoton", "isElectron", "isMuon"; four trajectory displacement variable: "part_D0", "part_ErrD0", "part_DZ", and "part_ErrDZ". The dataset format is fully compatible with the weaver package. This is an extension of the "Top Quark Tagging Reference Dataset". The generation cuts are unchanged (see Pythia cards) but several simulation improvements are included in the generation pipeline. The updated detector card provides more accurate simulations and a larger set of jet features.