LHC_HH4b_mixeddata / README.md
alefisico's picture
Fix LaTeX delimiters: use \(...) for inline KaTeX and $$ block equations
6301c8e verified
|
Raw History Blame Contribute Delete
14.5 kB
metadata
annotations_creators:
  - machine-generated
language_creators:
  - found
language:
  - en
license: cc0-1.0
multilinguality:
  - monolingual
size_categories:
  - 10M<n<100M
task_categories:
  - tabular-classification
  - tabular-regression
task_ids: []
pretty_name: >-
  CMS Open Data HH4b Reconstructed, Mixed and Simulated Data for Multijet
  Background Modelling
tags:
  - physics
  - high-energy-physics
  - particle-physics
  - lhc
  - cern
  - cms
  - open-data
  - di-higgs
  - multijet-background
  - background-estimation
  - hemisphere-mixing
  - domain-adaptation
  - density-ratio-estimation
  - tabular
configs:
  - config_name: data_3b
    data_files: data/3b/*.parquet
  - config_name: mixeddata_4b
    data_files: mixeddata_all/4b/*.parquet
  - config_name: ttbar_3b
    data_files: ttbar/3b/*.parquet
  - config_name: ttbar_4b
    data_files: ttbar/4b/*.parquet
  - config_name: GluGluToHHTo4B_cHHH1
    data_files: GluGluToHHTo4B_cHHH1/*.parquet
  - config_name: ZH4b
    data_files: ZH4b/*.parquet
  - config_name: ZZ4b
    data_files: ZZ4b/*.parquet

CMS Open Data: HH → 4b Multijet Background Modelling and Hemisphere-Mixed Dataset

DOI License: CC0-1.0 CERN Open Data Portal Open In Colab Jupyter Notebook


🚀 Quickstart Tutorial Notebook

An interactive end-to-end tutorial is provided in notebooks/quickstart.ipynb:

  • 1-Click Run in Colab: Click the Open In Colab badge above to launch and execute directly in your browser without downloading 14.5 GB locally.
  • Hands-on Demonstrations: Fast HTTP Parquet loading with pandas/polars, inspecting feature categories across simulation (623 cols), collision data (276 cols), and synthetic mixed data (191 cols), plotting jet kinematic distributions ((p_T, \eta, \phi)), visualizing DeepJet (b)-tagging domain shifts, and training a machine learning reweighting baseline.

1. Overview & Physics Challenge

This dataset contains flat tabular (Apache Parquet) collision data and simulated physics processes released by the CMS Collaboration on the CERN Open Data Portal (Record cms-93965).

It is prepared for machine-learning-based background modelling, density ratio estimation, and domain-adaptation reweighting in multijet searches for di-Higgs production in the four-bottom-quark final state:

pp→HH→(bbˉ)(bbˉ) \mathrm{pp} \to HH \to (b\bar{b})(b\bar{b})

recorded with the CMS detector at the CERN Large Hadron Collider (LHC) during Run 2 (2016–2018) at center-of-mass energy (\sqrt{s} = 13\text{ TeV}).

The Background Challenge

In the all-hadronic 4b final state, the dominant background arises from non-resonant quantum chromodynamic (QCD) multijet production. First-principles Monte Carlo simulations cannot model non-resonant multijet events with sufficient precision, and leading-order cross sections are computationally intractable. Consequently, particle physicists employ data-driven methods by transferring kinematic properties from an orthogonal, high-statistics source domain (three-tag events, 3b) into the target domain (four-tag events, 4b).

Furthermore, collision data in the signal-sensitive mass window (the Signal Region, SR) must remain strictly blinded to avoid confirmation bias. To provide an authentic, signal-free benchmark with an unblinded ground truth, this dataset provides synthetic 4b collision data generated via hemisphere mixing.


2. Machine Learning Task Formulation (Domain Adaptation & Density Ratio Estimation)

This benchmark presents a canonical, real-world machine learning problem: high-dimensional density ratio extrapolation under domain shift.

flowchart LR
    subgraph Source_Domain ["Source Domain: Sideband (SB)"]
        D3b_SB["3b Collision Data (SB)"]
        D4b_SB["4b Mixed Pseudo-Data (SB)"]
    end

    subgraph Learning ["Model Learning"]
        direction TB
        Reweighter["Train Reweight Classifier w_θ(x)<br>p(x|4b) / p(x|3b)"]
    end

    subgraph Target_Domain ["Target Domain: Signal Region (SR)"]
        D3b_SR["3b Collision Data (SR)"]
        Pred_SR["Predicted 4b Background (SR)"]
        GroundTruth_SR["Replica 0 Mixed Ground Truth (SR)"]
    end

    D3b_SB --> Reweighter
    D4b_SB --> Reweighter
    Reweighter -.->|"Extrapolate w_θ"| D3b_SR
    D3b_SR --> Pred_SR
    Pred_SR <-->|"Evaluate Transfer Error"| GroundTruth_SR

Formal Problem Definition

Let (\mathbf{x} \in \mathcal{X} \subset \mathbb{R}^d) denote an event feature vector. The feature space is partitioned into two disjoint kinematic regions:

  1. Source Domain ((\mathcal{X}_{\mathrm{src}})): The Sideband (SB) ring.
  2. Target Domain ((\mathcal{X}_{\mathrm{tgt}})): The Signal Region (SR) disk centered at candidate Higgs masses ((m_{H_1}, m_{H_2}) \approx (125, 120)\text{ GeV}).

Participants have access to:

  • Control data: (\mathcal{D}{3b}) available in both (\mathcal{X}{\mathrm{src}}) and (\mathcal{X}_{\mathrm{tgt}}).
  • Target data in the Source Domain: (\mathcal{D}_{4b}^{\mathrm{src}}) (both real collision data and synthetic mixed data).
  • Resonant background simulation: (\mathcal{D}_{t\bar{t}}) in both domains (subtracted to isolate the QCD multijet continuum).
  • Downstream signal template: (\mathcal{D}{HH}) in (\mathcal{X}{\mathrm{tgt}}).

The Objective: Learn a density ratio operator (w_\theta(\mathbf{x}) = \frac{p(\mathbf{x} \mid 4b)}{p(\mathbf{x} \mid 3b)}) strictly using data in (\mathcal{X}{\mathrm{src}}), and apply it to control data in (\mathcal{X}{\mathrm{tgt}}) to predict the target distribution:

p^4b(x∣x∈Xtgt)=wθ(x)⋅p3b(x∣x∈Xtgt) \hat{p}_{4b}(\mathbf{x} \mid \mathbf{x} \in \mathcal{X}_{\mathrm{tgt}}) = w_\theta(\mathbf{x}) \cdot p_{3b}(\mathbf{x} \mid \mathbf{x} \in \mathcal{X}_{\mathrm{tgt}})

The predicted distribution is evaluated against the unblinded synthetic ground truth (Replica 0) in the Signal Region.


3. Dataset Splits and Statistics

The dataset totals approximately 14.54 GB across 202 Parquet files and up to 623 feature columns.

Split / Config Name Data Origin Kinematic Domain Event Count Features Description & ML Workflow Role
data_3b Real Collisions Source (SB) + Target (SR) ~2.57M 276 High-statistics source domain (3b) to be reweighted into 4b.
mixeddata_4b Hemisphere Mixed Source (SB) + Target (SR) ~2.99M 191 Synthetic 4b collision data. For standard benchmark comparisons, use 1 replica (e.g. Replica 0) as the target real data.
ttbar_3b Simulation Source (SB) + Target (SR) ~1.20M 623 Top-quark pair production ((t\bar{t})) in the 3b category (subtracted during reweighting).
ttbar_4b Simulation Source (SB) + Target (SR) ~130k 623 Top-quark pair production ((t\bar{t})) in the 4b category.
GluGluToHHTo4B_cHHH1 Simulation Target (SR) ~150k 623 Standard Model di-Higgs signal template ((\kappa_\lambda = 1)) for sensitivity benchmarking.
ZH4b Simulation Target (SR) ~100k 623 Standard Model Electroweak candle ((ZH \to b\bar{b}b\bar{b})).
ZZ4b Simulation Target (SR) ~100k 623 Standard Model Electroweak candle ((ZZ \to b\bar{b}b\bar{b})).

4. Synthetic Data via Hemisphere Mixing & Benchmark Protocol

Evaluating density estimation or domain adaptation on real physical collision data faces a fundamental hurdle: events in the target Signal Region (SR) must remain blinded.

To overcome this, synthetic 4b events are generated using empirical hemisphere mixing:

  1. Spatial Bisection: Each collision event is partitioned into two spatial hemispheres ((H^+) and (H^-)) across the plane orthogonal to the transverse thrust axis (\hat{\mathbf{t}}).
  2. Empirical Library Query: An empirical library containing hundreds of thousands of hemispheres is constructed from real 4b collision data.
  3. Nearest-Neighbor Recombination: Each hemisphere of a base event is independently replaced by its closest kinematic match queried from the library using a (k)-d tree under a normalized Mahalanobis metric.
flowchart TD
    A["Observed 4b Collision Event"] --> B["Compute Transverse Thrust Axis"]
    B --> C["Bisect into Hemispheres H+ and H-"]
    C --> D["Query Nearest-Neighbor from 4b Library (k-d Tree)"]
    D --> E["Recombine Matches into Synthetic 4b Event"]
    E --> F["Resonant Signals Destroyed<br>Fidelity & Detector Effects Preserved"]

📌 Benchmark Guideline: Using 1 Mixed Replica as Real Data

  • For benchmark evaluations, domain-adaptation comparisons, and model scoring, users should use only ONE of the mixed data samples (e.g., Replica 0, files ending with *_0.parquet) as the pseudo-data target representing real collision data.
  • The dataset contains 16 statistically independent replicas in total, but standard model evaluation and reporting should treat a single replica (e.g., Replica 0) as the observed dataset.

5. Feature Schema & Observable Dictionary

The columnar schema organizes detector measurements, composite invariant masses, and event weights:

📊 Expand Detailed Observable Dictionary and Units
Observable Category Sample Feature Names Physical Units Description
Leading Jet Kinematics Jet1_pt, Jet2_pt, Jet3_pt, Jet4_pt GeV Transverse momentum of the 4 leading reconstructed jets ((p_T > 40\text{ GeV})).
Jet1_eta, Jet2_eta, Jet3_eta, Jet4_eta Dimensionless Pseudorapidity (\eta = -\ln\tan(\theta/2)) ((
Jet1_phi, Jet2_phi, Jet3_phi, Jet4_phi Radians Azimuthal angle (\phi \in [-\pi, \pi]).
Jet1_mass, Jet2_mass, Jet3_mass, Jet4_mass GeV Reconstructed jet invariant mass.
CMS Flavour Tagging Jet1_btagDeepFlavB ... Jet4_btagDeepFlavB [0, 1] CMS DeepJet neural network probability that the jet originated from a bottom quark ((b)).
Jet1_btagDeepFlavQG ... Jet4_btagDeepFlavQG [0, 1] Quark-versus-gluon discriminator score.
Composite Jet Pairs m4j GeV Invariant mass of the four-jet system.
Candidate masses m_H1, m_H2 GeV Invariant mass of the two candidate reconstructed Higgs bosons chosen by minimizing pairing asymmetry.
Global Shape Observables nJet, nSelJets Integer Total reconstructed jet multiplicity and high-purity selected jets.
ptt GeV Scalar transverse thrust momentum.
Blinding & Partitions SR Boolean True if the event falls in the Signal Region disk.
SB Boolean True if the event falls in the Sideband ring.
threeTag, fourTag Boolean True for 3-tag and 4-tag event selections, respectively.
passHLT Boolean Trigger selection flag (High-Level Trigger passing state).
Event Weights weight Float Monte Carlo generator and cross-section weight.
pseudoTagWeight Float Weight accounting for 3-tag to 4-tag combinatoric sampling.
FvT Float Baseline Four-vs-Three-tag neural network reweighting factor.

6. Quickstart / Usage Examples

Loading Direct Parquet Partitions with Pandas or Polars

import pandas as pd

BASE_URL = "https://huggingface.co/datasets/alefisico/LHC_HH4b_mixeddata/resolve/main"

# Read sample partition directly in under a second
df_signal = pd.read_parquet(f"{BASE_URL}/GluGluToHHTo4B_cHHH1/part_UL17_0.parquet")
df_3b = pd.read_parquet(f"{BASE_URL}/data/3b/part_UL16_postVFP_F.parquet")

# For mixed data, load Replica 0 as the target pseudo-data
df_mixed_rep0 = pd.read_parquet(f"{BASE_URL}/mixeddata_all/4b/part_UL16_postVFP_F_0.parquet")

Loading via Hugging Face datasets

from datasets import load_dataset

# Stream a configuration
ds_stream = load_dataset("alefisico/LHC_HH4b_mixeddata", "mixeddata_4b", split="train", streaming=True)
sample_rows = list(ds_stream.take(100))

# Or load specific partition files directly
ds_file = load_dataset(
    "alefisico/LHC_HH4b_mixeddata", 
    data_files="mixeddata_all/4b/part_UL16_postVFP_F_0.parquet", 
    split="train"
)

7. Citation and Acknowledgements

This dataset is distributed under the Creative Commons CC0 1.0 Universal Public Domain Dedication.

When using this dataset in research or benchmarking, please cite the official CERN Open Data record:

@misc{cms_opendata_hh4b_mixeddata,
  author       = {{CMS Collaboration}},
  title        = {{HH4b reconstructed, mixed and simulated data in Parquet format for multijet background modelling}},
  year         = 2026,
  publisher    = {CERN Open Data Portal},
  doi          = {10.7483/OPENDATA.CMS.WWHU.D4ON},
  url          = {https://opendata.cern.ch/record/cms-93965}
}

For the machine learning reweighting framework and hemisphere mixing methodology, please refer to:

  • CMS Collaboration, Search for nonresonant Higgs boson pair production in the four b jet final state in proton-proton collisions at (\sqrt{s} = 13\text{ TeV}), Phys. Rev. Lett. 129, 081802 (2022).
  • CMS Collaboration, Search for the standard model Higgs boson produced in association with a W or Z boson and decaying into a pair of b quarks, Eur. Phys. J. C 81, 178 (2021).

8. Maintainers & Community

  • Maintainer: Alejandro Gómez Espinosa (CERN / Carnegie Mellon University)
  • Feedback & Questions: Please open an issue or thread in the Community tab of this repository!