DCTR: Pythia e+e- -> Z -> dijets datasets

A collection of datasets used in Neural Networks for Full Phase-space Reweighting and Parameter Tuning. Sample code for reproducing the results is available on GitHub.

Each dataset was generated with the Pythia 8.230 event generator. Particle-level \(e^+ e^- \to Z \to \text{dijet}\) events with about 100 particles in each event are clustered into jets using the anti-kt clustering algorithm (R = 0.8) with Fastjet 3.0.3. Every jet is presented as a list of constituents \((p_T, \eta, \phi, \text{particle ID}, \theta)\) where \(\theta = (\texttt{TimeShower:alphaSvalue}, \texttt{StringZ:aLund }, \texttt{StringFlav:probStoUD})\).

Training Datasets:
Each training file contains two arrays, X and Y.
X is an array of jets and Y is 0 (1) if the jet was generated with default (non-default) Pythia parameters. For a non-default jet (Y=1), the \(\theta\) in each constituent represents the value of the Pythia parameter that was used. Note that for a default jet (Y=0) the \(\theta\) in each constituent is not the default Pythia parameters, but \(\theta\) uniformly sampled in the same range as the Y=1 jets. 

The parameters were uniformly sampled in 

• \(\texttt{TimeShower:alphaSvalue} \in [0.10, 0.18]\)
• \(\texttt{StringZ:aLund } \in [0.50, 0.90]\)
• \(\texttt{StringFlav:probStoUD } \in [0.10, 0.30]\)

The 1D datasets are labeled by which parameter was changed, and the 3D dataset simultaneously vary all three parameters. 

Test Datasets:
Each test dataset consists of an dictionary containing: 

• 'jet': the jet constituents
• 'multiplicity': Number of particles in jet
• 'tau21': Nsubjettiness observable
• 'tau32': Nsubjettiness observable
• 'ECF_N3_B4': Energy Correlation Function(N=3, \(\beta\)=4)
• 'ECF_N4_B4': Energy Correlation Function(N=4, \(\beta\)=4)

The corresponding \(\theta\) values for each test set are described in the paper.