Skip to content

Benchmarks & datasets

Shared datasets are what make claims in this field comparable. Listed here are the ones used widely enough that a new method is expected to report on them. Datasets tied to a single topic are listed in that section instead.

Jet tagging & representations

  • JetClass, Qu, Li, Qian (arXiv:2202.03772) — a large-scale jet tagging dataset released with Particle Transformer; now the default benchmark for architecture comparisons.
  • Aspen Open Jets, (arXiv:2412.10504) — jets built from CMS Open Data, assembled for pre-training foundation models.
  • RODEM Jet Datasets, Zoch et al. (arXiv:2408.11616) — jet datasets aimed at method development and comparison.

Anomaly detection

  • LHC Olympics 2020, Kasieczka et al. (arXiv:2101.08320) — R&D and black-box datasets, with a summary paper comparing many submitted methods.
  • Dark Machines Anomaly Score Challenge, Aarrestad et al. (arXiv:2105.14027) — complementary in signals and event representation.
  • 40 MHz unsupervised new-physics dataset, Govorkova et al. (arXiv:2107.02157) — poses the problem under trigger-level constraints.

Simulation

  • CaloChallenge 2022, Krause et al. (arXiv:2410.21611) — three calorimeter datasets of increasing granularity, with agreed metrics; the reference benchmark for generative surrogates.
  • step2point, Zaborowska et al. (arXiv:2509.22340) — detailed shower simulation for data-representation studies.
  • ColliderML, Elitez et al. (arXiv:2512.15230) — an OpenDataDetector dataset at HL-LHC scale.

Uncertainty & inference

  • FAIR Universe HiggsML Uncertainty Challenge, Bhimji et al. (arXiv:2410.02867) — poses systematic-aware inference as a competition task.

Deployment

  • wa-hls4ml, Hawks et al. (arXiv:2511.05615) — benchmark and surrogate models for predicting FPGA resource use and latency.

Note

This page is curated, not exhaustive. If a dataset has become a standard reference point in your subfield, please propose it via a pull request.