jet-universe/jetclass2

Dataset Card for JetClass-II

JetClass-II is a large-scale and comprehensive dataset covering extensive large-radius jet signatures and a wide range of jet $p_{\mathrm{T}}p\_\backslash mathrm\{T\}$ and mass values.

The dataset is designed to develop large and comprehensive jet models, intended for various applications, to support extensive physics searches and measurements at the Large Hadron Collider (LHC) and beyond.

Dataset Details

Dataset Description

The dataset consists of three major parts based on the jet origin and its substructure:

1. Res2P: Generic $XX$ → 2 prong resonant jets.
2. Res34P: Generic $XX$ → 3 or 4 prong resonant jets.
3. QCD: Jets from QCD multijet background.

Each part is further subdivided into detailed categories, indicating which partons, leptons, or combinations thereof initiated the jet.

The three major parts (Res2P, Res34P, and QCD) are separately packed and can be downloaded individually for ease of use. The sizes of the training sets are 20M, 86M, and 28M entries, respectively. The dataset also includes validation and test sets, with the sizes for training/validation/test following a 4:1:1 ratio.

Every 100k entries (jets) are stored in a Parquet file. A complete view of the JetClass-II data files is shown in the table below.

Type	File name range	File number	total entries
Res2P, train	Res2P_0000.parquet—Res2P_0199.parquet	200	20M
Res2P, val	Res2P_0200.parquet—Res2P_0249.parquet	50	5M
Res2P, test	Res2P_0250.parquet—Res2P_0299.parquet	50	5M
Res34P, train	Res34P_0000.parquet—Res34P_0859.parquet	860	86M
Res34P, val	Res34P_0860.parquet—Res34P_1074.parquet	215	21.5M
Res34P, test	Res34P_1075.parquet—Res34P_1289.parquet	215	21.5M
QCD, train	QCD_0000.parquet—QCD_0279.parquet	280	28M
QCD, val	QCD_0280.parquet—QCD_0349.parquet	70	7M
QCD, test	QCD_0350.parquet—QCD_0419.parquet	70	7M

• License: MIT

Dataset Demo

Use [Colab] to inspect and visualize data in JetClass-II.

This demo will showcase visualizations of jets, annotated with the top 5 probability scores as interpreted by the Sophon model, the first application on JetClass-II.

Dataset Downloads

To facilitate downloading, the HTTP links for all data files are provided in filelist.txt.

Dataset Sources

• Repository: https://github.com/jet-universe/sophon
• Paper: https://arxiv.org/abs/2405.12972
• Demo: https://colab.research.google.com/github/jet-universe/sophon/blob/main/notebooks/Interacting_with_JetClassII_and_Sophon.ipynb

Uses

1. This dataset can be used to train models for various jet-related tasks, such as jet classification, jet property regression, and jet generation or reconstruction.
2. The dataset's extensive phase space coverage and high statistics enable model developers to focus on specific regions of interest, or work with the entire dataset, enabling the creation of specialized models for particular phase spaces or pre-training a more general model.
3. The dataset contains detailed low-level information to support customized model training strategies, including kinematic features, particle IDs, and trajectory displacement information for both jet constituent particles and relevant generator-level particles (see details in the next section).

Dataset Structure

The JetClass-II dataset includes the following variables:

1. part_*: Features for jet constituent particles (i.e., E-flow objects in Delphes).
2. jet_*: Features for jets. A specific variable is jet_label, which indicates the label in 188 classes.
3. genpart_*: Features for generator-level jet (GEN-jet) constituent particles. The GEN-jet is clustered from the stable particles generated by Pythia, excluding neutrinos, using the same clustering configuration. The GEN-jets are matched with jets based on the pseudoangular separation $\mathrm{\Delta }R\backslash Delta\; R$. Jets, ordered by decreasing $p_{\mathrm{T}}p\_\backslash mathrm\{T\}$, are paired with the closest unmatched GEN-jet. If no matched GEN-jet is found, the entry is left empty, which occurs in only 0.2–0.8% of cases.
4. genjet_*: Jet-level features for the matched GEN-jet.
5. aux_genpart_*: Auxiliary variables storing features of selected truth particles. Five types of particles are chosen if they are valid:
   1. The initial resonance $XX$ (in both 2-prong and 3/4-prong resonance cases).
   2. The two secondary resonances $YY$ produced by $XX$ ( $X\to Y_1{Y}_{2}X\; \backslash to\; Y\_1Y\_2$ ) in the 3/4-prong resonance case.
   3. The direct decay products (partons and leptons) from $XX$ and $YY$.
   4. The subsequent decay products of tau leptons in case (iii).
   5. The partons ( $p_{\mathrm{T}}p\_\backslash mathrm\{T\}$ > 5 GeV) matched within a QCD jet.

Variable	Type	Description	Exists in JetClass?
For jet constituent particles
part_px	vector<float>	particle's $p_{x}p\_x$	✔️
part_py	vector<float>	particle's $p_{y}p\_y$	✔️
part_pz	vector<float>	particle's $p_{z}p\_z$	✔️
part_energy	vector<float>	particle's energy	✔️
part_deta	vector<float>	difference in pseudorapidity $\eta \backslash eta$ between the particle and the jet axis	✔️
part_dphi	vector<float>	difference in azimuthal angle $\varphi \backslash phi$ between the particle and the jet axis	✔️
part_d0val	vector<float>	particle's transverse impact parameter value $d_{0}d\_0$, in mm	✔️
part_d0err	vector<float>	error of the particle's transverse impact parameter ${\sigma }_{d_0}\backslash sigma\_\{d\_0\}$, in mm	✔️
part_dzval	vector<float>	particle's longitudinal impact parameter value $d_{z}d\_z$, in mm	✔️
part_dzerr	vector<float>	error of the particle's longitudinal impact parameter ${\sigma }_{d_z}\backslash sigma\_\{d\_z\}$, in mm	✔️
part_charge	vector<int32_t>	particle's electric charge	✔️
part_isElectron	vector<bool>	if the particle is an electron (abs(pid)==11)	✔️
part_isMuon	vector<bool>	if the particle is an muon (abs(pid)==13)	✔️
part_isPhoton	vector<bool>	if the particle is an photon (pid==22)	✔️
part_isChargedHadron	vector<bool>	if the particle is a charged hadron (charge!=0 && !isElectron && !isMuon)	✔️
part_isNeutralHadron	vector<bool>	if the particle is a neutral hadron (charge==0 && !isPhoton)	✔️
For jets
jet_pt	float	jet's transverse momentum $p_{\mathrm{T}}p\_\backslash mathrm\{T\}$	✔️
jet_eta	float	jet's pseudorapidity $\eta \backslash eta$	✔️
jet_phi	float	jet's azimuthal angle $\varphi \backslash phi$	✔️
jet_energy	float	jet's energy	✔️
jet_sdmass	float	jet's soft-drop mass	✔️
jet_nparticles	int32_t	number of jet constituent particles	✔️
jet_tau1	float	jet's $NN$-subjettiness variable ${\tau }_1\backslash tau\_1$	✔️
jet_tau2	float	jet's $NN$-subjettiness variable ${\tau }_2\backslash tau\_2$	✔️
jet_tau3	float	jet's $NN$-subjettiness variable ${\tau }_3\backslash tau\_3$	✔️
jet_tau4	float	jet's $NN$-subjettiness variable ${\tau }_4\backslash tau\_4$	✔️
jet_label	int32_t	jet's label index in JetClass-II, detailed in the above table	🆕
For GEN-jet constituent particles (if a GEN-jet is found matched to a jet)
genpart_px	vector<float>	particle's $p_{x}p\_x$	🆕
genpart_py	vector<float>	particle's $p_{y}p\_y$	🆕
genpart_pz	vector<float>	particle's $p_{z}p\_z$	🆕
genpart_energy	vector<float>	particle's energy	🆕
genpart_jet_deta	vector<float>	difference in pseudorapidity $\eta \backslash eta$ between the particle and the jet (not the GEN-jet) axis	🆕
genpart_jet_dphi	vector<float>	difference in azimuthal angle $\varphi \backslash phi$ between the particle and the jet (not the GEN-jet) axis	🆕
genpart_x	vector<float>	$xx$ coordinate of the particle's production vertex, in mm	🆕
genpart_y	vector<float>	$yy$ coordinate of the particle's production vertex, in mm	🆕
genpart_z	vector<float>	$zz$ coordinate of the particle's production vertex, in mm	🆕
genpart_t	vector<float>	$tt$ coordinate of the particle's production vertex, in mm/c	🆕
genpart_pid	vector<int32_t>	particle's PDGID	🆕
For GEN-jets (if matched to a jet)
genjet_pt	float	GEN-jet's transverse momentum $p_{\mathrm{T}}p\_\backslash mathrm\{T\}$	🆕
genjet_eta	float	GEN-jet's pseudorapidity $\eta \backslash eta$	🆕
genjet_phi	float	GEN-jet's azimuthal angle $\varphi \backslash phi$	🆕
genjet_energy	float	GEN-jet's energy	🆕
genjet_sdmass	float	GEN-jet's soft-drop mass	🆕
genjet_nparticles	int32_t	number of GEN-jet constituent particles	🆕
For selected truth particles
aux_genpart_pt	vector<float>	selected truth particles' $p_{\mathrm{T}}p\_\backslash mathrm\{T\}$	✔️ (different rules to select truth particles)
aux_genpart_eta	vector<float>	selected truth particles' $\eta \backslash eta$	✔️ (different rules to select truth particles)
aux_genpart_phi	vector<float>	selected truth particles' $\varphi \backslash phi$	✔️ (different rules to select truth particles)
aux_genpart_mass	vector<float>	selected truth particles' mass	✔️ (different rules to select truth particles)
aux_genpart_pid	vector<int32_t>	selected truth particles' PDGID	🆕
aux_genpart_isResX	vector<bool>	if the particle is the initial resonance $XX$	🆕
aux_genpart_isResY	vector<bool>	if the particle is the secondary resonance $YY$	🆕
aux_genpart_isResDecayProd	vector<bool>	if the particle is the direct decay product (parton and lepton) from $XX$ and $YY$	🆕
aux_genpart_isTauDecayProd	vector<bool>	if the particle is the subsequent decay product of tau leptons	🆕
aux_genpart_isQcdParton	vector<bool>	if the particle is the parton with $p_{\mathrm{T}}p\_\backslash mathrm\{T\}$ > 5 GeV stored in the QCD jet case	🆕

Dataset Creation

The dataset is generated using MadGraph + Pythia + Delphes.

During the Delphes (fast simulation) step, the pileup (PU) effect, with an average of 50 PU interactions, is emulated to mimic the realistic LHC collision environment. The PUPPI algorithm is then applied to remove the PU, correcting the E-flow objects used to cluster jets. This distinguishes it from the original JetClass dataset. The Delphes card can be found in the jetclass2-generation repository.

The complete generation script (the one-stop MadGraph + Pythia + Delphes production) and the n-tuplizer script are provided in the jetclass2-generation repository to facilitate reproducibility.

Citation

If you use the JetClass-II dataset, please cite:

BibTeX:

@article{Li:2024htp,
    author = "Li, Congqiao and Agapitos, Antonios and Drews, Jovin and Duarte, Javier and Fu, Dawei and Gao, Leyun and Kansal, Raghav and Kasieczka, Gregor and Moureaux, Louis and Qu, Huilin and Suarez, Cristina Mantilla and Li, Qiang",
    title = "{Accelerating Resonance Searches via Signature-Oriented Pre-training}",
    eprint = "2405.12972",
    archivePrefix = "arXiv",
    primaryClass = "hep-ph",
    month = "5",
    year = "2024"
}

@dataset{jetclass2,
    author        = "Li, Congqiao and Agapitos, Antonios and Drews, Jovin and Duarte, Javier and Fu, Dawei and Gao, Leyun and Kansal, Raghav and Kasieczka, Gregor and Moureaux, Louis and Qu, Huilin and Suarez, Cristina Mantilla and Li, Qiang",
    collaboration = "Jet Universe",
    title         = "JetClass-II",
    year          = "2025",
    url           = "https://huggingface.co/datasets/jet-universe/jetclass2",
    doi           = "10.57967/hf/6336",
    publisher     = "Hugging Face"
}

Glossary

A good resource is the CERN Open Data Portal Glossary: https://opendata.cern.ch/search?q=&f=type%3AGlossary&l=list&order=asc&p=1&s=10&sort=title

Jet: A jet is a shower of hadrons, which originate from quarks or gluons, clustered together after being produced in particle collisions. A large-radius jet is clustered using a larger radius parameter $RR$ (0.8 in this dataset) and may result from a collection of nearby quarks, gluons, and other particles.

Constituent particle: The particles (reconstructed hadrons, electrons, muons, or photons) that form the jet after clustering.

GEN-jet: A generator-level jet, reconstructed from a list of stable truth particles in a simulation.

Truth particle: Particles produced during a collision in a simulation. Initial truth particles are directly generated in the hard collision process, but they may undergo decays, intermediate emissions, and parton showering to produce the stable particles ultimately observed by the detector. These stable particles are used to cluster GEN-jets.

Pseudorapidity $\eta \backslash eta$: The pseudorapidity $\eta \backslash eta$ is a coordinate that describes the angle of a particle (or jet) produced in an event relative to the beam axis. It is calculated as $\eta =-\ln (\tan \frac{\theta }{2})\backslash eta\; =\; -\; \backslash ln\; \backslash left\; (\; \backslash tan\; \backslash frac\{\backslash theta\}\{2\}\; \backslash right\; )$, with $\theta \backslash theta$ the angle between the three-momentum and the beam axis. $\eta =0\backslash eta=0$ means the produced particle/jet is perpendicular to the beam axis, while a higher pseudorapidity means it is closer to the beam.

Transverse momentum $p_{\mathrm{T}}p\_\backslash mathrm\{T\}$: The component of the momentum of a particle (or jet) that is transverse (i.e., perpendicular) to the beam axis.

Pseudoangular distance $\mathrm{\Delta }R\backslash Delta\; R$: $\mathrm{\Delta }R(a,b)=\sqrt{({\eta }_a-{\eta }_b{)}^2+({\varphi }_a-{\varphi }_b{)}^2}\backslash Delta\; R(a,\backslash ,b)\; =\; \backslash sqrt\{(\backslash eta\_a\; -\; \backslash eta\_b)^2\; +\; (\backslash phi\_a\; -\; \backslash phi\_b)^2\}$, where $\eta (\varphi )\backslash eta\backslash ;(\backslash phi)$ is the pseudorapidity (azimuthal angle) of the momentum of a particle or a jet. A particle is considered matched to the jet if $\mathrm{\Delta }R(\mathrm{p}\mathrm{a}\mathrm{r}\mathrm{t}\mathrm{i}\mathrm{c}\mathrm{l}\mathrm{e},\mathrm{j}\mathrm{e}\mathrm{t}\mathrm{a}\mathrm{x}\mathrm{i}\mathrm{s})>R_0\backslash Delta\; R\backslash mathrm\{(particle,\backslash ;jet\backslash ,axis)\}\; >\; R\_0$, where $R_{0}R\_0$ is the jet radius parameter.

Impact parameter: The distance of the closest approach of the track to the collision point.

Particle's PDGID: A unique identifier assigned to each particle type by the Particle Data Group (PDG). The full PDGID table can be accessed here.

[More Information Needed]

Dataset Card Authors and Contacts

@jmgduarte, @colizz