StarEmbed/ZTF_40k

StarEmbed Benchmark Data set

The StarEmbed benchmark data set is available here in huggingface datasets format.

The ~40,000 multi-band ZTF light curves are available at StarEmbed/data. The train/validation/test splits are included alongside the additional anom split used for the OOD benchmarking.

Available columns

Each star in data set has the following fields:

Top-level columns

Column	Type	Description
sourceid	string	Catalina Surveys DR1 (CSDR1) source id of the variable star (e.g. CSS_J082956.4-044426)
bands_data	dict / struct (keys g, i, r)	per-band light curves; each band is a struct (or null if absent) holding the arrays in the table below.
period	float64	catalog period of the variable star, in days (range ≈ 0.13 – 885)
class_str	string	variable-star class label; one of EW, EA, RRab, RRc, RRd, RS CVn, LPV
ra	float64	right ascension (J2000) in decimal degrees
dec	float64	declination (J2000) in decimal degrees

Per-band fields inside each band of bands_data (g / r / i)

Field	Type	Description
target	list of floats	magnitude, measure of brightness (AB system) — the raw light curve
past_feat_dynamic_real	list of floats	mag_err, 1σ uncertainty on magnitude, aligned with target
feat_dynamic_real	list of floats	delta_t, time gap (days) between each two consecutive observations
mjd	list of floats	observation time in Modified Julian Date, aligned with target
length	int64	number of observations in this band's light curve

All four splits (train, validation, test, anom) share the identical schema described above. The train/validation/test splits contain the seven in-distribution classes listed under class_str, whereas the anom split is an anomaly-detection holdout whose class_str values are a disjoint set of out-of-distribution classes — Beta_Lyrae, Blazhko, ACEP, Cep-II, HADS, LADS, ELL, Hump, PCEB, EA_UP — that never appear in training.

---

Citation

This work was accepted to the main track of ICML 2026 as well as the AI4Physics workshop. If you use or refer to this dataset please cite our work.

@article{StarEmbed,
       author = {{Li}, Weijian and {Chen}, Hong-Yu and {Rehemtulla}, Nabeel and {Shah}, Ved G. and {Wu}, Dennis and {Kim}, Dongho and {Lin}, Qinjie and {Miller}, Adam A. and {Liu}, Han},
        title = "{StarEmbed: Benchmarking Time Series Foundation Models on Astronomical Observations of Variable Stars}",
      journal = {arXiv e-prints},
     keywords = {Solar and Stellar Astrophysics, Instrumentation and Methods for Astrophysics, Artificial Intelligence},
         year = 2025,
        month = oct,
          eid = {arXiv:2510.06200},
        pages = {arXiv:2510.06200},
          doi = {10.48550/arXiv.2510.06200},
archivePrefix = {arXiv},
       eprint = {2510.06200},
 primaryClass = {astro-ph.SR},
       adsurl = {https://ui.adsabs.harvard.edu/abs/2025arXiv251006200L},
      adsnote = {Provided by the SAO/NASA Astrophysics Data System}
}

The StarEmbed data set is curated from ZTF observations and the Catalina Surveys Periodic Variable Star Catalog. If you use the StarEmbed data please also cite the works below.

@article{bellm2018zwicky,
  title={The Zwicky Transient Facility: system overview, performance, and first results},
  author={Bellm, Eric C and Kulkarni, Shrinivas R and Graham, Matthew J and Dekany, Richard and Smith, Roger M and Riddle, Reed and Masci, Frank J and Helou, George and Prince, Thomas A and Adams, Scott M and others},
  journal={Publications of the Astronomical Society of the Pacific},
  volume={131},
  number={995},
  pages={018002},
  year={2018},
  publisher={IOP Publishing}
}

@article{drake2014catalina,
  title={The catalina surveys periodic variable star catalog},
  author={Drake, AJ and Graham, MJ and Djorgovski, SG and Catelan, M and Mahabal, AA and Torrealba, G and Garc{\'\i}a-{\'A}lvarez, D and Donalek, C and Prieto, JL and Williams, R and others},
  journal={The Astrophysical Journal Supplement Series},
  volume={213},
  number={1},
  pages={9},
  year={2014},
  publisher={IOP Publishing}
}025}
}