README.md

---
pretty_name: Delphes ROOT Samples
tags:
  - physics
  - high-energy-physics
  - root
  - delphes
---

# Delphes ROOT Samples

This dataset contains detector-level ROOT samples for high-energy-physics studies of Higgs bosons, top quarks, photons, and related processes in proton-proton collisions. The core simulated samples were produced for pretraining and downstream evaluation of machine-learning models in collider physics.

Events were generated with **MadGraph5_aMC@NLO 2.7.3**, showered and hadronized with **Pythia 8.235**, and passed through **Delphes 3.4.2** with an ATLAS-like detector configuration. Unless noted otherwise, the collision energy is √s = 13 TeV.

The repository also contains analysis-specific and derived samples. A process is not intrinsically a signal or background sample: that role depends on the analysis and is not encoded by the directory layout.

## Dataset description

### Pretraining samples

The core pretraining collection spans 12 Standard Model processes covering major Higgs-boson and top-quark production modes:

- Higgs production: gluon fusion (ggF), vector-boson fusion (VBF), associated W and Z production (WH, ZH), top-pair-associated production (tt̄H), and single-top-associated production (tHq; represented in filenames as `tHjb`).
- Top production: single-top, tt̄, tt̄γγ, tt̄W, three-top, and four-top (tt̄tt̄) production.

For inclusive samples, Higgs bosons and top quarks are allowed to decay inclusively. The primary campaign contains approximately 10 million events per process, except tHq and four-top production, for which approximately 5 million events per process were generated. Individual repository files may be subsets, variants, or derived products; `metadata/samples.csv` is the canonical file inventory.

### Evaluation and downstream samples

Additional samples support transfer-learning and generalization studies beyond the pretraining processes. They include Standard Model tt̄H, H→γγ; a CP-odd top-Yukawa tt̄H, H→γγ variant based on the Higgs Characterization model; stop-pair production in the MSSM; and flavor-changing neutral-current (FCNC) top-quark processes.

For the stop sample, pp → t̃₂ t̃₂-bar, with m(t̃₂) = 582 GeV and m(t̃₁) = 400 GeV.

with each heavier stop decaying to a lighter stop and a Higgs boson. The FCNC sample is based on tt̄ production in which one top quark undergoes a flavor-changing decay involving a Higgs boson and a light quark.

## Event generation and detector simulation

The main simulation chain is **MadGraph5_aMC@NLO → Pythia 8 → Delphes.**

| Component | Version / configuration |
| --- | --- |
| Matrix-element generation | MadGraph5_aMC@NLO 2.7.3 |
| Parton shower / hadronization | Pythia 8.235 |
| Heavy-particle decays | Pythia 8.235 |
| Top-quark and W-boson decays | MadSpin |
| Detector simulation | Delphes 3.4.2 |
| Detector configuration | ATLAS-like |
| Collision system and energy | Proton-proton, √s = 13 TeV |
| Jet algorithm | anti-kₜ, R = 0.4 |

Core pretraining samples were generated at next-to-leading order (NLO) in QCD. Other generation or analysis qualifiers are encoded in filenames where applicable and should be checked per file.

## Reconstructed physics objects

Detector-level objects use ATLAS-like selections. Photons require pT ≥ 20 GeV and |η| ≤ 2.37; electrons require pT ≥ 10 GeV and |η| ≤ 2.47. For both, the electromagnetic-calorimeter transition region 1.37 < |η| < 1.52 is excluded. Muons require pT ≥ 10 GeV and |η| ≤ 2.7.

Jets are reconstructed with anti-kₜ, R = 0.4, and selected with pT ≥ 25 GeV and |η| ≤ 2.5. Jets within ΔR < 0.4 of a selected photon or lepton are removed to avoid double counting, where ΔR = √((Δη)² + (Δφ)²). b jets are identified by matching reconstructed jets to b quarks within ΔR = 0.4, with efficiency corrections intended to approximate the corresponding ATLAS b-tagging performance.

## Repository layout

```text
samples/
  higgs/              ggF, VBF, VH, ttH, and tHjb samples
  top/                single-top, ttbar, ttV, multi-top, FCNC, and stop samples
  photon/             diphoton and tt+diphoton samples

derived/
  multilabel/         samples augmented for multilabel workflows

analyses/
  hww/                H→WW analysis samples
  th-cp/              tH CP-analysis samples
  thbb/               tH→bb analysis samples

metadata/
  samples.csv         machine-readable file inventory
```

Some directory and filename labels retain the naming convention of the original production campaign and may differ slightly from publication notation.

## Relationship to the associated paper

The associated paper documents a specific pretraining and evaluation study; this repository is a broader collection of Delphes and analysis-derived samples. **The presence of a file in this repository does not imply that it was used in the paper's pretraining, downstream evaluation, or reported results.**

| Collection | Relationship to the paper dataset | Contents |
| --- | --- | --- |
| `samples/higgs/`, `samples/top/`, `samples/photon/` | General-purpose physics samples; includes the core processes described in the pretraining collection and additional process/channel variants. | Higgs, top-quark, FCNC, stop, diphoton, and photon-associated production samples. |
| `derived/multilabel/` | Derived data products, rather than independent event-generation campaigns. | Samples augmented for multilabel workflows. |
| `analyses/th-cp/` | Analysis-specific data products. | Dedicated tH CP-analysis samples. |
| `analyses/hww/` | Analysis-specific data products not covered by the paper's pretraining benchmark. | H→WW analysis ntuples, including signal and background variants. |
| `analyses/thbb/` | Analysis-specific data products not covered by the paper's pretraining benchmark. | tH→bb̄ samples with CP-angle and top-reconstruction variants. |

For reproducibility, use the paths named in the publication or its accompanying analysis materials to identify the exact samples for a reported result. Do not infer study membership solely from a directory name or filename qualifier.

## File-level metadata

`metadata/samples.csv` provides the canonical distributed-file inventory, including paths and byte sizes. Where available, users should also record or obtain the event count, cross section, sum of event weights, generator configuration, decay and filter selections, and the sample's study role (pretraining, downstream evaluation, analysis-specific, or other). These properties can vary across files, including files with similar names.

## Filename conventions

Common qualifiers describe individual files, not necessarily every sample:

- `LO`, `NLO`: leading- or next-to-leading-order generation.
- `inc`: inclusive sample; `truth`: includes truth-level information.
- `1l`, `2l`, `had`, `lep`, `loose`, `tight`: channel or selection variants.
- `multilabel`: derived sample for multilabel workflows.

Consult `metadata/samples.csv` for the canonical path and byte size of each ROOT file.

## Data format and loading

Samples are distributed as ROOT files containing Delphes event records. ROOT is not natively supported by the Hugging Face Dataset Viewer; use [ROOT](https://root.cern/), [uproot](https://uproot.readthedocs.io/), or [Awkward Array](https://awkward-array.org/). Trees and branches can differ among production and analysis samples, so inspect each target file rather than assuming a common schema.

```python
from huggingface_hub import hf_hub_download
import uproot

path = hf_hub_download(
    repo_id="HWresearch/Delphes",
    repo_type="dataset",
    filename="samples/higgs/ggf/ggF_NLO_inc.root",
)

with uproot.open(path) as root_file:
    print(root_file.keys())
    tree = root_file["Delphes"]
    print(tree.keys())
```

## ATLAS Open Data evaluation

The associated study also evaluates generalization with external ATLAS Open Data at 13 TeV from 2015–2016 proton-proton collisions (integrated luminosity ≈ 36 fb⁻¹). These data provide a domain shift relative to the Delphes samples because they use the ATLAS detector simulation and reconstruction chain.

The GamGam collection requires at least two photons with pT ≥ 25 GeV and is enriched in H→γγ processes. The 1LMET30 collection requires at least one lepton with pT ≥ 7 GeV and missing transverse energy ≥ 30 GeV; the associated analysis uses WWW, ZZZ, WWZ, and WZZ triboson processes.

The ATLAS triboson evaluation samples use Sherpa 2.2.2 and the ATLAS Geant4-based simulation and reconstruction chain. They therefore test generalization across unseen processes, generators, shower configurations, detector-simulation frameworks, and reconstruction pipelines. **These ATLAS Open Data are external evaluation data and are not redistributed in this repository.** Obtain them from the official ATLAS Open Data release and follow its terms of use.

## Intended uses

This dataset is intended for research in collider physics and machine learning, including representation learning, pretraining and transfer learning, event classification, anomaly and out-of-distribution detection, Higgs and top-quark studies, CP-sensitive observables, rare or BSM processes, and robustness studies across simulation and reconstruction domain shifts.

## Limitations

- **Simulation-to-data differences:** Delphes is a fast, approximate detector simulation and does not reproduce the complete ATLAS simulation, reconstruction, calibration, triggering, pileup treatment, or data-quality conditions. Results on these samples alone should not be interpreted as expected performance on ATLAS collision data.
- **Generator dependence:** Models may learn generator-, shower-, detector-, or reconstruction-specific features. Evaluation with alternative generators or detector simulations is recommended.
- **Sample sizes and rates:** Generated counts do not encode physical relative rates. Phenomenological analyses should apply suitable cross sections, branching fractions, efficiencies, luminosity normalization, and event weights.
- **Analysis-specific selections:** Some files use selections or representations for particular analyses; inspect filenames, metadata, and ROOT branches before combining them.
- **Physics coverage:** The collection is diverse but is not an exhaustive simulation of Standard Model proton-proton collisions.

## Citation

If you use this dataset in scientific work, please cite the associated open-access publication, [*Pretrained event classification model for high energy physics analysis*](https://doi.org/10.1088/1748-0221/21/08/P08006), and the relevant event-generation, showering, detector-simulation, and physics-model references. The preprint is also available on [arXiv:2412.10665](https://arxiv.org/abs/2412.10665).

```bibtex
@article{Ho_2026,
  author  = {Ho, Joshua and Roberts, Ryan and Han, Shuo and Wang, Haichen},
  title   = {Pretrained event classification model for high energy physics analysis},
  journal = {Journal of Instrumentation},
  volume  = {21},
  number  = {08},
  pages   = {P08006},
  year    = {2026},
  doi     = {10.1088/1748-0221/21/08/P08006},
  url     = {https://arxiv.org/abs/2412.10665}
}
```

## Dataset repository

Hugging Face dataset repository: `HWresearch/Delphes`.