HH4b reconstructed, mixed and simulated data in Parquet format for multijet background modelling

Description

Flat tabular (Apache Parquet) datasets prepared for machine-learning-based background modelling and kinematic reweighting in multijet searches for Higgs boson pair production in the four-bottom-quark final state ($HH \rightarrow b\bar{b}b\bar{b}$) with the CMS detector at the LHC.

Modelling the dominant QCD multijet background is a primary challenge in this final state, since simulation does not describe QCD multijet events with sufficient accuracy, and data-driven background templates are used instead. The record contains:

• Real collision data: CMS Run 2 proton-proton collision data at 13 TeV. Only three-tag (3b) events are released; four-tag (4b) collision data are not included.
• Mixed data (mixeddata_all): synthetic background events built by combining independent hemispheres taken from different events. This breaks event-level correlations (eliminating signal contamination) while preserving detector acceptance and single-hemisphere kinematics.
• Simulated signal and backgrounds: Monte Carlo samples for gluon-gluon fusion $HH$ signal, top-quark pair production ($t\bar{t}$), $ZH$ and $ZZ$.

All events are pre-selected to fall in either the Signal Region (SR) or the Sideband (SB), i.e. SR | SB == True. Each event carries 622 columns: event identifiers, all reconstructed jet variables from the parent picoAOD files (padded to 12 jets), Higgs-candidate classifier inputs, background-modelling weights (JCM and FvT) and the outputs of a signal-vs-background multi-classifier.

Dataset characteristics

How were these data selected?

Events are preselected by requiring at least four AK4 jets with $p_T > 40$ GeV and $|\eta| < 2.4$ that pass the tight jet identification and the pileup jet identification, and that are separated from isolated leptons. Jets are b-tagged with the DeepJet algorithm. Events with at least four jets passing the medium working point form the four-tag sample. Events with exactly three jets passing the loose working point form the three-tag sample, which is used to model the multijet background from data. Because the three-tag events contain exactly three loose b-tagged jets, the two samples do not overlap.

More information about the selection can be found in their respective publications:

• Eur. Phys. J. C 84 (2024) 712
• CMS-HIG-24-010

3b / 4b splitting. The mixeddata_all/ and ttbar/ directories are split into 3b/ and 4b/ subdirectories according to the threeTag and fourTag flags. For data/, only the 3b/ subdirectory is released: four-tag collision data are not part of this record. All mixeddata_all events fall in 4b/, as expected since the hemisphere-mixing method is designed to model the four-tag background. The GluGluToHHTo4B_cHHH1/, ZH4b/ and ZZ4b/ directories are not split.

Merged $t\bar{t}$. The dileptonic (TTTo2L2Nu), semileptonic (TTToSemiLeptonic) and fully hadronic (TTToHadronic) samples are merged into ttbar/. A string column process records the original sample of each event, and output file names are prefixed with the original sample name (e.g. part_TTToHadronic_part_UL18.parquet). Generator weights were already normalised in the inputs, so no cross-section rescaling was applied.

How were these data validated?

• Mutual exclusivity: threeTag and fourTag are never both true for the same event in any split dataset.
• Event conservation: for every input file, the sum of events in the 3b/ and 4b/ outputs equals the original event count.
• Untouched samples: GluGluToHHTo4B_cHHH1, ZH4b and ZZ4b have identical event counts before and after restructuring.

How can you use these data?

The flat Parquet files contain event parameters, classifier inputs, background weights, and reconstructed jet features.

Event Identifiers & Reconstructed Jets (picoAOD)

Standard CMS event parameters and all reconstructed jet variables. We store ALL Jet_ variables present in the parent picoAOD files, padded up to 12 jets per event:

• run: Run number.
• luminosityBlock: Luminosity block number.
• event: Event number.
• nJet: Total number of reconstructed jets in the event.

For each jet index $i \in [1, 12]$, the following variables are stored (padded with -1.0, 0, or NaN as appropriate):

• Kinematics:
  • Jet{i}_pt: Transverse momentum ($p_T$).
  • Jet{i}_eta: Pseudorapidity ($\eta$).
  • Jet{i}_phi: Azimuthal angle ($\phi$).
  • Jet{i}_mass: Mass.
• B-tagging Scores:
  • Jet{i}_btagDeepFlavB: DeepJet b-tagging score.
  • Jet{i}_btagDeepB: DeepCSV b-tagging score.
  • Jet{i}_btagCSVV2: CSVv2 b-tagging score.
• Energy Fractions & Constituents:
  • Jet{i}_chHEF / Jet{i}_neHEF: Charged / Neutral Hadron Energy Fractions.
  • Jet{i}_chEmEF / Jet{i}_neEmEF: Charged / Neutral Electromagnetic Energy Fractions.
  • Jet{i}_muEF: Muon Energy Fraction.
  • Jet{i}_nConstituents: Number of jet constituents.
• Jet Calibration & Quality:
  • Jet{i}_area: Active area of the jet.
  • Jet{i}_rawFactor: Raw energy correction scale factor.
  • Jet{i}_bRegCorr: B-jet energy regression correction factor.
  • Jet{i}_jetId: Jet ID quality flag (integer indicating tight/loose quality).
  • Jet{i}_puId: Pileup jet ID discriminator.
  • Jet{i}_cleanmask: Cleaning mask flag.
• Generator-Level Matching (Simulation only):
  • Jet{i}_genJetIdx: Index matching to the generator-level jet collection.
  • Jet{i}_hadronFlavour / Jet{i}_partonFlavour: Hadron / Parton flavour categorization.

Classifier Input & Selection (HCR Input Friend)

Variables describing candidate Higgs jets (candidate jets) vs. remaining jets (non-candidate jets), along with region definition flags:

Candidate Jets (CanJet — padded to 4)

• CanJet{i}_pt / CanJet{i}_eta / CanJet{i}_phi / CanJet{i}_mass: Kinematics of the 4 jets selected as Higgs decay candidates ($i \in [1, 4]$).
• CanJet{i}_btagScore: DeepJet b-tagging score of candidate jets.
• CanJet{i}_area: Jet active area.
• CanJet{i}_rawFactor: Raw energy correction factor.

Non-Candidate Jets (NotCanJet — padded to 8)

• NotCanJet{i}_pt / NotCanJet{i}_eta / NotCanJet{i}_phi / NotCanJet{i}_mass: Kinematics of the remaining jets not selected for Higgs candidates ($i \in [1, 8]$).
• NotCanJet{i}_isSelJet: Boolean flag indicating if the non-candidate jet passes analysis selections.

Region & Event Selection Flags

• SR: Signal Region flag (True if event falls in the mass signal window).
• SB: Sideband Region flag (True if event falls in the sideband window).
• ZHSR, ZZSR, HHSR: Specific Signal Region flags for ZH, ZZ, and HH channels.
• threeTag / fourTag: Flags indicating if the event has exactly 3 or 4 b-tagged jets.
• passHLT: Flag indicating if the event passes the HLT trigger requirements.
• nSelJets: Count of selected jets in the event.
• xW, xbW: $\chi^2$ metrics for the top mass and W boson reconstruction hypotheses.
• weight: Generator-level event weight.

Background Modeling Weights (JCM & FvT Friends)

Re-weighting factors used to model backgrounds for the 4-tag signal region:

• pseudoTagWeight: Event-level Jet Combinatorics Method (JCM) weight to model multijet background using 3-tag data.
• FvT: Four-vs-Three tag reweighting factor (FvT classifier output) to correct for kinematical differences between 3-tag and 4-tag events.

Signal vs. Background Classifier Output (SvB MA Friend)

Output variables from the Signal vs. Background (SvB) multi-classifier (SpanNet):

• pmj: Multijet background probability.
• ptt: $t\bar{t}$ background probability.
• pzz: $ZZ$ signal probability.
• pzh: $ZH$ signal probability.
• phh: $HH$ signal probability.
• ps: Probability score discriminator for background multijet.
• passMinPs: Boolean flag representing passing the minimum background probability cut.
• zz / zh / hh: Likelihood ratios for the respective signal hypotheses.
• q_1234, q_1324, q_1423: Probabilities for different quadjet pairing permutations.
• tt_vs_mj: Discriminator between $t\bar{t}$ and QCD multijet backgrounds.
• ps_zz, ps_zh, ps_hh: Targeted probability discriminators for ZZ, ZH, and HH signals.

How to Load and Read the Data

The files can be read with any Parquet reader and do not need CMS-specific software. Example with pandas:

Files and indexes

Disclaimer

These open data are released under the Creative Commons Zero v1.0 Universal license.

Neither the experiment(s) ( CMS ) nor CERN endorse any works, scientific or otherwise, produced using these data.

This release has a unique DOI that you are requested to cite in any applications or publications.