Update README.md

README.md

@@ -85,3 +85,88 @@ configs:
   - split: anom
     path: data/anom-*
 ---
+
+# StarEmbed Benchmark Data set
+
+The StarEmbed benchmark data set is available here in huggingface datasets format.
+
+The ~40,000 multi-band ZTF light curve 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`    | int64                             | source/object ID of the variable star (e.g. `102311274853951453`)                                    |
+| `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.|
+| `item_id`     | string                            | unique series id, `{sourceid}_{band}` for the primary band (usually `g`), e.g. `102311274853951453_g` |
+| `start`       | pandas.Timestamp (timestamp[us])  | date of first observation (first `mjd` converted to a timestamp)                                     |
+| `freq`        | string                            | sampling frequency (e.g. `"1D"` for daily)                                                           |
+| `period`      | float64                           | catalog period of the variable star, in days (range ≈ 0.13 – 885)                                    |
+| `objid`       | string                            | object identifier, same as `item_id` (`{sourceid}_{band}`)                                           |
+| `csdr1_id`    | string                            | Catalina Surveys DR1 (CSDR1) cross-match id, e.g. `CSS_J082956.4-044426`                             |
+| `class_str`   | string                            | variable-star class label; one of `EW`, `EA`, `RRab`, `RRc`, `RRd`, `RS CVn`, `LPV`                  |
+
+**Per-band fields inside each band of `bands_data` (`g` / `r`)**
+| 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 (encoded as numeric strings '3', '7', '9'–'17') 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}
+}
+```