Training data for "The Well-Tempered Likelihood" — Pythia 8.317 e+e- jet observables

Training dataset for the DCTR (Deep neural networks using Classification for Tuning and Reweighting) surrogate likelihood used in "The Well-Tempered Likelihood: Honest Confidence Intervals for Misspecified Models." The file contains 2 million jets from simulated e+e- → Z → hadrons events at √s = 91.2 GeV, generated with Pythia 8.317 (Monash tune). Final-state particles are clustered with the e+e- anti-kT algorithm (ee_genkt with p = −1, R = 0.8) via FastJet, using E-scheme recombination with jets ordered by energy. File contents (pythia_dctr_train_2M.npz): Y — class labels (0 = reference, 1 = varied parameters), shape (2000000,) params — (alphaS, aLund, probStoUD) per jet, shape (2000000, 3) jet_mass — jet mass from E-scheme four-vector sum multiplicity — particle multiplicity number_of_kaons — number of charged kaons (|PID| = 321) ED — energy dispersion √(∑ Eᵢ²) / ∑ Eᵢ width — energy-weighted mean 3D opening angle to the E-scheme jet axis, ∑ Eᵢ θᵢ / ∑ Eᵢ Parameter ranges (uniform prior for Y=1 jets): αs (TimeShower:alphaSvalue): [0.10, 0.18] aLund (StringZ:aLund): [0.50, 0.90] probStoUD (StringFlav:probStoUD): [0.10, 0.30] Reference point (Y=0 jets): (αs, aLund, probStoUD) = (0.1365, 0.68, 0.217) Keywords: machine learning, jet substructure, likelihood inference, simulation-based inference, goodness of fit, DCTR, Pythia License: CC-BY-4.0 Related identifiers: Supersedes Zenodo record 3518708 (original dataset with pp anti-kT clustering) GitHub: https://github.com/bnachman/well-tempered-likelihood