OneAstroBench

Official v0 evaluation subsets for OneAstroBench (astronomical foundation-model benchmark).

Configs (tasks)

Config	Task	Split(s)	Notes
OA-A1-gz10-morph-v0	GZ10 morphology (10-class)	test	images (5,96,96) + label
OA-A2-object-class-v0	STAR / GALAXY / QSO	test	images (4,96,96) + label
OA-B1-sspp-stellar-v0	SSPP stellar params	test	1D spectra + Teff/logg/[Fe/H]
OA-B2-provabgs-v0	ProVABGS galaxy params	test	images + log_mstar/ssfr/z
OA-C1-lens-retrieval-v0	Strong-lens retrieval	query, gallery	+ qrels.parquet
OA-E1-img-phot2spec-v0	Image+phot → spectrum	test	image, photometry, target spectrum

Arrays are stored as flat float32 lists (image, spectrum_*) plus an explicit *_shape column so the Dataset Viewer shows collapsible numeric lists (not Base64 unknown blobs).

from datasets import load_dataset
import numpy as np
ds = load_dataset("xiaofanloveme/OneAstroBench", "OA-A1-gz10-morph-v0", split="test")
row = ds[0]
img = np.asarray(row["image"], dtype=np.float32).reshape(row["image_shape"])

Format notes

• HF-compatible: yes — standard Parquet + dataset card YAML configs.
• Prefer OneAstroBench TaskDataset loaders for evaluation.
• This release is the frozen v0 eval subset (seed=42 where sampling applies), not the full multi-hundred-GB source archives.

Citation

If you use OneAstroBench, please cite the project repository and the original survey catalogs (Legacy Survey, DESI, SDSS/SSPP, Galaxy Zoo, etc.).