EiffL/desi_x_ls_aion

AION-1 embeddings of DESI × Legacy Survey galaxies

95,895 galaxies, each carrying token-level AION-1 embeddings of its Legacy Survey image and DESI spectrum (embedded separately), together with its PROVABGS physical properties (redshift, stellar mass, metallicity, age, star-formation rate, …).

The embeddings are ready-made features for downstream tasks — similarity search, property regression, anomaly detection — with no GPU or raw-pixel/spectrum processing required.

How it was built

1. Cross-match — three catalogs from the Multimodal Universe HATS release were joined at 1″ using LSDB, with PROVABGS as the anchor:

   Role	Catalog	Modality
   anchor	UniverseTBD/mmu_desi_provabgs	spectroscopic galaxy properties
   spectra	UniverseTBD/mmu_desi_edr_sv3	DESI EDR SV3 spectra
   imaging	hugging-science/mmu_legacysurvey_dr10_south_21	Legacy Survey DR10 south (DECaLS) 4-band cutouts

2. Tokenize + embed — each modality was encoded with AION-1 base (Parker et al. 2025, arXiv:2510.17960), Polymathic AI's multimodal astronomical foundation model:

   • the 4-band (g,r,i,z) 160×160 image cutout → LegacySurveyImage codec → AION encoder → 576 × 768 embedding
   • the DESI spectrum (flux, ivar, mask, λ) → DESISpectrum codec → AION encoder → 273 × 768 embedding

   The full per-token embedding sequences are stored (no pooling), so you can mean-pool, attend, or select tokens as needed. Inference ran in float32; embeddings are stored as float16 (values ≤ ~75 in magnitude, well within float16 range).

Dataset structure

477 parquet shards under data/. Columns:

Column	Type	Description
embedding_image	float16 [576, 768]	AION-1 token embeddings of the Legacy Survey image
embedding_spectrum	float16 [273, 768]	AION-1 token embeddings of the DESI spectrum
object_id_provabgs	string	DESI TARGETID
ra_provabgs, dec_provabgs	float64	sky position (deg)
Z_HP_provabgs	float32	redshift
LOG_MSTAR_provabgs	float32	log₁₀ stellar mass
Z_MW_provabgs	float32	mass-weighted metallicity
TAGE_MW_provabgs	float32	mass-weighted age (Gyr)
AVG_SFR_provabgs	float32	average star-formation rate
+ 11 more	float32	PROVABGS magnitudes, weights, and quality columns

PROVABGS properties come from the DESI PROVABGS value-added catalog (Hahn et al. 2023, arXiv:2208.08512).

Usage

The full dataset is ~117 GB — stream it:

import numpy as np
from datasets import load_dataset

ds = load_dataset("EiffL/desi_x_ls_aion", streaming=True, split="train")

row = next(iter(ds))
im_emb = np.array(row["embedding_image"], dtype=np.float32)   # (576, 768)
sp_emb = np.array(row["embedding_spectrum"], dtype=np.float32)  # (273, 768)

# single vector per galaxy (e.g. for similarity search / linear probes)
im_vec = im_emb.mean(axis=0)  # (768,)

print(row["Z_HP_provabgs"], row["LOG_MSTAR_provabgs"])

References

• AION-1: Parker et al. 2025, arXiv:2510.17960 — model · code
• The Multimodal Universe: MMU Collaboration 2024, arXiv:2412.02527 — HATS release blog
• PROVABGS: Hahn et al. 2023, arXiv:2208.08512

License follows the source Multimodal Universe catalogs: CC-BY-4.0.