HiggsML Uncertainty Challenge Public Dataset

This dataset has been created for the HiggsML Uncertainty Challenge, a NeurIPS 2024 competition. A detailed documentation is available in the challenge white paper

The tabular dataset is created using the particle physics simulation tools Pythia 8.2 and Delphes 3.5.0. The proton-proton collision events are generated with a center of mass energy of 13 TeV using Pythia8. Subsequently, these events undergo the Delphes tool to produce simulated detector measurements. We used an ATLAS-like detector description to make the dataset closer to experimental data. The events are divided into two groups:

1. Higgs boson signal (H→ττ)
2. ZZ boson background (Z→ττ)
3. Diboson background (VV→ττ)
4. ttbar background (ttˉ)

⚠️ Note: The "LHC events" is the average number in this category in a pseudo-experiment corresponding to running of the Large Hadron Collider for 10 fb−1, corresponding to approximately 800 billion inelastic proton collisions, or 2 weeks in summer 2024 conditions

Higgs Signal:

The Higgs bosons are produced with all possible production modes and decay into two tau leptons. The tau leptons are further allowed to decay into all possible final states, but only final states with one lepton (electron or muon) and one hadron tau decay are kept.

Z boson Background:

Only background events coming from Z bosons are included in this challenge. While simulating the process, interference effects between Z bosons and photons are included. Similar to signal events, only the tau-tau decay mode of the Z boson is included in the dataset.

⚠️ Note:

The training events have weights.

Event Weights:

Event weights are defined as:

w=Cross-Section × LuminosityTotal number of generated eventsw=Total number of generated eventsCross-Section × Luminosity

The challenge is considering a scenario of analyzing proton-proton collision data of 10 fb−1 luminosity collected by the ATLAS experiment.

Features in the data

Prefix-less variables

Weight, Label, DetailedLabel, have a special role and should NOT be used as regular features for the model:

Primary Features

The variables prefixed with PRI (for PRImitives) are “raw” quantities about the bunch collision as measured by the detector, essentially parameters of the momenta of particles.

Derived Features

These variables are derived from the primary variables with the help of derived_quantities.py. 

Preselection Cuts

⚠️ Note: The post-selection cuts are the cuts made after systematics is applied.
⚠️ Note: The Dataset might not be properly shuffled. 
One could use dataset.py from the dataset repository (see below). 

Utility Software

Alongside the dataset, a GitHub repository with the relevant code for reading and analysing it is made available. This includes a Jupyter notebook starting kit, simple baseline models, and code to run the challenge and generate the score. The repository also has a sample dataset, a subset of the main dataset, to let users experience the challenge software without downloading the much larger dataset.

The code for dataset generation is provided in a dedicated repository: https://github.com/FAIR-Universe/genHEPdata. This repository also contains a Dockerfile, which facilitates the installation of the necessary software dependencies.