Rakib/PhyGHT-TopQuark

PhyGHT Top Quark Events

Lossless Parquet redistribution of the official PhyGHT top-quark pair simulation dataset, published by Mohammed Rakib. Each configuration has 10,000 events: mu200 models extreme HL-LHC pileup, and mu60 the lower pileup scenario. These are simulation events, not measured detector collisions.

Load

from datasets import load_dataset
events = load_dataset("Rakib/PhyGHT-TopQuark", "mu200")
event = events["test"][0]
assert len(events["train"]) == 8000

Use "mu60" for the other configuration. streaming=True permits streaming without loading the entire configuration. Explicit train/validation/test split paths are in this card's metadata; no executable dataset script is required.

Schema and units

Each row is one original simulated event. event_id is its zero-based index in the corresponding official pickle; unique identity is (pileup, event_id). pileup is 200 or 60. All original rows, column ordering, variable-length lists, float64 values, and int64 associations are preserved without quantization.

• jets: list of jets, each [pT, eta, phi, mass, Efrac, Mfrac] (float64).
• trks: list of tracks, each [pT, eta, phi, mass, d0, z0, track_id, truth_label] (float64 as in the source, including integral identifiers and labels).
• jet_trk_IDs: one list of int64 track identifiers for each jet, in original order. These refer to track column 6, not positions in the event track list.
• The model uses the first four jet columns and the first six track columns, converts inputs to float32, and predicts the last two jet columns. Track auxiliary labels are truth_label == -1 for hard-scatter signal; other labels denote pileup. No truth labels or identifiers enter the feature encoders.

The official walkthrough explicitly labels track pT in GeV. Eta and the fraction targets are dimensionless; phi is an azimuthal angle. The release does not assign unverified units to mass, d0, or z0. Consult the paper and simulation producers before making assumptions about unverified units.

Reproducible event splits

The official preprocessing algorithm is preserved exactly: numpy.random.RandomState(42).permutation(10000), then the first 8,000 events for train, next 1,000 for validation, and last 1,000 for test. Shard row order matches these indices. Indices for each configuration are saved in splits/. All three splits are disjoint within a configuration. Event numbers across different pileup configurations do not establish independent physical provenance or alignment beyond the source simulation's own guarantees.

Integrity and validation

Official source files were size- and MD5-verified before deserialization. Every converted event was compared to the original at full numerical precision; every event also produced bit-identical float32 model features, targets, track labels, and jet-to-track edges through the original dataset implementation. mu200_validation.json and mu60_validation.json contain the measured checks, raw hashes, multiplicities, and any association identifiers not present in the global track list. The original code ignores such absent identifiers when constructing edges; this redistribution retains them in the raw associations. manifest.json provides SHA256 hashes for release files.

License, provenance, and limitations

The Zenodo API explicitly specifies CC BY 4.0 for the dataset. Attribution to the six authors and the original DOI is required. The associated software has a separate MIT license. Source metadata is included verbatim for inspection. Simulation-to-detector transfer, other processes and pileup conditions, and unverified physical units require independent validation. Event lists can be large; the Dataset Viewer may be less convenient than streaming Python access.

Citation and sources

• Paper: PhyGHT, KDD 2026
• Official code: https://github.com/rAIson-Lab/PhyGHT
• Original dataset: https://doi.org/10.5281/zenodo.18746388

@misc{rakib2026phyghtphysicsguidedhypergraphtransformer,
  title={PhyGHT: Physics-Guided HyperGraph Transformer for Signal Purification at the HL-LHC},
  author={Mohammed Rakib and Luke Vaughan and Shivang Patel and Flera Rizatdinova and Alexander Khanov and Atriya Sen},
  year={2026}, eprint={2602.20475}, archivePrefix={arXiv}, primaryClass={hep-ex},
  url={https://arxiv.org/abs/2602.20475}
}