kshitijd/astro-multimodal-570k

Astro Multimodal 570k: Pre-Cross-Matched Astronomical Dataset

The first publicly available, pre-cross-matched astronomical dataset that unifies three modality types -- spectra, light curves, and images -- into a single table. ~570k objects across three populations (stars, galaxies, AGN) are joined from 12 major surveys into ready-to-use rows: no cross-matching required.

Why This Dataset?

Existing multimodal astronomical datasets either provide raw survey collections that users must cross-match themselves, or cover only 1-2 modalities for a single population:

Dataset	Objects	Modalities	Surveys	Populations	Pre-joined?
This dataset	570k	Spectra + Light Curves + Images	12	Stars, Galaxies, AGN	Yes
Multimodal Universe	100M+ (separate)	Spectra + LC + Images	20+	Mixed	No (raw collections)
AstroCLIP	198k	Spectra + Images	2	Galaxies only	Yes
AstroM3	21k	Spectra + LC + Metadata	6	Variable stars only	Yes
DESI/HSC	19k	Spectra + Images	2	Galaxies only	Yes

This dataset is ready for multimodal representation learning, transfer learning across wavelengths, population classification, and any task that benefits from having multiple views of the same astronomical object in a single row.

---

Quick Start

Installation

pip install datasets numpy pandas pyarrow
# Optional for visualization:
pip install matplotlib astropy

System requirements: Loading a single shard (~5000 rows) needs ~2 GB RAM. Loading the full dataset needs ~150-200 GB RAM. For large-scale work, stream or load shards individually (see below).

Load and explore

from datasets import load_dataset
import numpy as np

# Stream without downloading everything
ds = load_dataset("kshitijd/astro-multimodal-570k", streaming=True)

# Or download fully
ds = load_dataset("kshitijd/astro-multimodal-570k")

Get a star with an infrared spectrum

row = ds["train"][0]  # or iterate with streaming
if row["population"] == "star" and row["apogee_flux"] is not None:
    flux = np.array(row["apogee_flux"], dtype=np.float32)  # (7514,) normalized IR spectrum
    flux_err = np.array(row["apogee_flux_err"], dtype=np.float32)

Get a galaxy with UV + IR images

for row in ds["train"]:
    if row["population"] == "galaxy" and row["galex_fuv"] is not None and row["unwise_w1"] is not None:
        fuv = np.array(row["galex_fuv"], dtype=np.float32).reshape(64, 64)   # GALEX far-UV
        w1 = np.array(row["unwise_w1"], dtype=np.float32).reshape(64, 64)    # WISE 3.4 micron
        break

Plot a spectrum

import matplotlib.pyplot as plt
import numpy as np

row = ds["train"][0]
if row["apogee_flux"] is not None:
    flux = np.array(row["apogee_flux"], dtype=np.float32)
    # APOGEE wavelength grid: 3 detectors, 7514 good pixels
    # Approximate wavelength range: 1.51-1.70 microns
    plt.figure(figsize=(12, 3))
    plt.plot(flux, lw=0.5)
    plt.xlabel("Pixel")
    plt.ylabel("Normalized Flux")
    plt.title(f"APOGEE Spectrum: {row['object_id']}")
    plt.tight_layout()
    plt.savefig("spectrum.png", dpi=150)

Plot a light curve

if row["ztf_time"] is not None:
    time = np.array(row["ztf_time"])
    mag = np.array(row["ztf_mag"])
    magerr = np.array(row["ztf_magerr"])
    band = np.array(row["ztf_band"])

    plt.figure(figsize=(10, 4))
    for b in np.unique(band):
        mask = band == b
        plt.errorbar(time[mask], mag[mask], yerr=magerr[mask], fmt='.', label=f"Band {b}", ms=3)
    plt.gca().invert_yaxis()
    plt.xlabel("HJD")
    plt.ylabel("Magnitude")
    plt.legend()
    plt.title(f"ZTF Light Curve: {row['object_id']}")
    plt.tight_layout()
    plt.savefig("lightcurve.png", dpi=150)

Plot image cutouts across wavelengths

fig, axes = plt.subplots(1, 5, figsize=(15, 3))
bands = [("galex_fuv", "GALEX FUV"), ("legacy_g", "Legacy g"),
         ("twomass_j", "2MASS J"), ("unwise_w1", "WISE W1"), ("unwise_w2", "WISE W2")]

for ax, (col, label) in zip(axes, bands):
    if row[col] is not None:
        img = np.array(row[col], dtype=np.float32).reshape(64, 64)
        ax.imshow(img, origin="lower", cmap="gray")
        ax.set_title(label)
    else:
        ax.text(0.5, 0.5, "No data", ha="center", va="center", transform=ax.transAxes)
    ax.axis("off")

plt.suptitle(f"{row['object_id']} ({row['population']})")
plt.tight_layout()
plt.savefig("cutouts.png", dpi=150)

Filter by population

# All stars with spectra AND images
stars_multimodal = ds["train"].filter(
    lambda x: x["population"] == "star" and x["n_spectra"] > 0 and x["n_images"] > 0
)

# AGN with light curves
agn_variable = ds["train"].filter(
    lambda x: x["population"] == "agn" and x["n_lightcurves"] > 0
)

Memory-efficient loading with pandas

import pandas as pd

# Load just one shard (~5000 rows, ~2 GB RAM)
df = pd.read_parquet("00000.parquet")

# Load only specific columns (much less RAM)
df = pd.read_parquet("00000.parquet", columns=["object_id", "population", "ra", "dec",
                                                 "apogee_flux", "n_spectra", "n_images"])

# Iterate over all shards without loading everything
import glob
for f in sorted(glob.glob("*.parquet")):
    chunk = pd.read_parquet(f)
    stars = chunk[chunk["population"] == "star"]
    # process stars...
    del chunk  # free memory

Build a PyTorch dataset

import torch
from torch.utils.data import Dataset, DataLoader
import pandas as pd
import numpy as np
import glob

class AstroDataset(Dataset):
    """Memory-efficient dataset that loads one shard at a time."""

    def __init__(self, shard_dir, population=None, require_modalities=None):
        self.files = sorted(glob.glob(f"{shard_dir}/*.parquet"))
        # Build index: (shard_idx, row_idx) for each valid object
        self.index = []
        for si, f in enumerate(self.files):
            df = pd.read_parquet(f, columns=["population", "n_spectra", "n_lightcurves", "n_images"])
            for ri in range(len(df)):
                if population and df.iloc[ri]["population"] != population:
                    continue
                if require_modalities:
                    if "spectra" in require_modalities and df.iloc[ri]["n_spectra"] == 0:
                        continue
                    if "images" in require_modalities and df.iloc[ri]["n_images"] == 0:
                        continue
                self.index.append((si, ri))
            del df

        self._cache_si = -1
        self._cache_df = None

    def __len__(self):
        return len(self.index)

    def __getitem__(self, idx):
        si, ri = self.index[idx]
        if si != self._cache_si:
            self._cache_df = pd.read_parquet(self.files[si])
            self._cache_si = si
        row = self._cache_df.iloc[ri]

        sample = {"object_id": row["object_id"], "population": row["population"]}

        # Spectrum
        if row.get("apogee_flux") is not None and isinstance(row["apogee_flux"], (list, np.ndarray)):
            sample["spectrum"] = torch.tensor(np.array(row["apogee_flux"], dtype=np.float32))

        # Image (example: 2MASS J-band)
        if row.get("twomass_j") is not None and isinstance(row["twomass_j"], (list, np.ndarray)):
            img = np.array(row["twomass_j"], dtype=np.float32).reshape(64, 64)
            sample["image"] = torch.tensor(img).unsqueeze(0)  # (1, 64, 64)

        return sample

# Usage:
dataset = AstroDataset("./shards/", population="star", require_modalities=["spectra", "images"])
loader = DataLoader(dataset, batch_size=32, shuffle=True, num_workers=2)

---

Dataset Summary

Population	Count	Spectra Coverage	Light Curve Coverage	Image Coverage
Stars	300k	99.7% (APOGEE)	23.2% (TESS + ZTF)	100% (2MASS + GALEX + unWISE)
Galaxies	200k	20.5% (SDSS + DESI)	--	96.3% (GALEX + unWISE)
AGN	100k	100% (SDSS)	8.5% (ZTF)	97.1% (GALEX + unWISE)

Multimodal Coverage

Population	>= 2 modality types	All 3 modality types
Stars	99.8%	23.1%
Galaxies	19.2%	0% (no light curves by design)
AGN	97.3%	8.4%

Per-Source Coverage Detail

Stars (300k)

Source	Column	Coverage
APOGEE DR17	apogee_flux	299,143 (99.7%)
Gaia BP/RP	flatiron_gaia_coeff	147,096 (49.0%)
GALAH DR4	galah_flux	22,123 (7.4%)
TESS	flatiron_tess_flux	41,823 (13.9%)
ZTF DR24	ztf_time	28,049 (9.3%)
2MASS	twomass_j/h/k	299,984+ (100%)
GALEX	galex_fuv/nuv	168k-224k (56-75%)
unWISE	unwise_w1/w2	22,729 (7.6%)

Galaxies (200k)

Source	Column	Coverage
SDSS	sdss_flux	18,839 (9.4%)
DESI (spectra)	flatiron_desi_spectrum_flux	24,697 (12.3%)
DESI (metadata)	flatiron_desi_z	150,263 (75.1%)
GALEX	galex_fuv/nuv	191k-193k (95-96%)
unWISE	unwise_w1/w2	24,986 (25.0%)

AGN (100k)

Source	Column	Coverage
SDSS	sdss_flux	99,995 (100%)
ZTF DR24	ztf_time	8,538 (8.5%)
GALEX	galex_fuv/nuv	92k-97k (92-97%)
unWISE	unwise_w1/w2	24,986 (25.0%)

---

Data Sources

Spectra

Source	Instrument	Wavelength	Resolution	Population	Coverage
APOGEE DR17	APOGEE (APO + LCO)	1.51--1.70 um (IR)	R ~ 22,500	Stars	299k / 300k
Gaia DR3 BP/RP	Gaia BP/RP	330--1050 nm	R ~ 50--100	Stars	147k / 300k
GALAH DR4	HERMES (AAT)	4713--7887 A	R ~ 28,000	Stars	22k / 300k
SDSS DR17	BOSS / eBOSS	3600--10400 A	R ~ 2000	Galaxies, AGN	119k / 300k
DESI EDR	DESI	3600--9800 A	R ~ 2000--5000	Galaxies, AGN	25k / 300k

Light Curves

Source	Instrument	Bandpass	Cadence	Population	Coverage
TESS	TESS	600--1000 nm	2--30 min	Stars, AGN	42k / 400k
ZTF DR24	ZTF (Palomar)	g, r, i	1--3 day	Stars, AGN	37k / 400k

Images

Source	Instrument	Bands	Pixel Scale	Cutout Size	Population	Coverage
2MASS	2MASS	J, H, K	1 arcsec/px	64 x 64	Stars	300k / 300k
GALEX	GALEX	FUV, NUV	1.5 arcsec/px	64 x 64	All	452k / 570k
unWISE	WISE	W1, W2	2.75 arcsec/px	64 x 64	All	48k / 570k
Legacy Survey	DECam / Mosaic / 90Prime	g, r, z	0.262 arcsec/px	64 x 64	All	4k / 570k

---

Schema

Core columns (all objects)

Column	Type	Description
object_id	string	Unique identifier (APOGEE 2MASS ID for stars, PROVABGS ID for galaxies, SDSS DR14Q name for AGN)
ra	float64	Right ascension (degrees, J2000)
dec	float64	Declination (degrees, J2000)
population	string	"star", "galaxy", or "agn"
n_spectra	int	Count of spectral datasets with data for this object
n_lightcurves	int	Count of light curve datasets with data
n_images	int	Count of image bands with data

Spectra columns

Column	Type	Shape	Description
apogee_flux	list[float32]	(7514,)	APOGEE normalized flux, cropped to good detector pixels
apogee_flux_err	list[float32]	(7514,)	APOGEE flux uncertainty
galah_flux	list[float32]	variable	GALAH combined 4-band flux
galah_lambda	list[float32]	variable	GALAH wavelength array (A)
sdss_flux	list[float32]	variable	SDSS/BOSS spectral flux (10^-17 erg/s/cm^2/A)
sdss_loglam	list[float32]	variable	SDSS log10(wavelength / A)
sdss_ivar	list[float32]	variable	SDSS inverse variance
flatiron_gaia_coeff	list[float32]	(110,)	Gaia BP/RP spectral coefficients
flatiron_desi_spectrum_flux	list[float32]	variable	DESI coadded spectral flux
flatiron_desi_spectrum_lambda	list[float32]	variable	DESI wavelength array (A)
flatiron_desi_spectrum_ivar	list[float32]	variable	DESI inverse variance

Light curve columns

Column	Type	Shape	Description
ztf_time	list[float64]	variable	ZTF observation times (HJD)
ztf_mag	list[float32]	variable	ZTF PSF magnitudes
ztf_magerr	list[float32]	variable	ZTF magnitude uncertainties
ztf_band	list	variable	ZTF filter code
flatiron_tess_time	list[float64]	variable	TESS observation times (BTJD)
flatiron_tess_flux	list[float32]	variable	TESS normalized flux
flatiron_tess_flux_err	list[float32]	variable	TESS flux uncertainty

Image columns

All image columns are stored as nested lists representing 64 x 64 pixel cutouts. Reconstruct with np.array(row["col"], dtype=np.float32).reshape(64, 64).

Column	Type	Description
twomass_j, twomass_h, twomass_k	list[list[float32]]	2MASS J/H/K-band cutout
galex_fuv, galex_nuv	list[list[float32]]	GALEX far-UV / near-UV cutout
unwise_w1, unwise_w2	list[list[float32]]	unWISE W1 (3.4 um) / W2 (4.6 um) cutout
legacy_g, legacy_r, legacy_z	list[list[float32]]	Legacy Survey g/r/z-band cutout

Metadata columns

The dataset includes ~300 metadata columns from source surveys, prefixed by survey name. Key examples:

Column	Description
apogee_teff	APOGEE effective temperature (K)
apogee_logg	APOGEE surface gravity (log g)
flatiron_gaia_phot_g_mean_mag	Gaia G-band apparent magnitude
flatiron_gaia_parallax	Gaia parallax (mas)
flatiron_gaia_teff_gspphot	Gaia photometric effective temperature
flatiron_desi_z	DESI spectroscopic redshift
flatiron_desi_spectype	DESI spectral classification
agn_redshift	AGN redshift from SDSS DR14Q

---

File Format and System Requirements

Format

Parquet shard files, up to 5000 rows each (~120 shards total). Populations are interleaved -- filter on population to select types. Total on-disk size: ~80 GB compressed.

System Requirements

Use Case	RAM	Disk	Notes
Stream one shard at a time	2 GB	1 GB	Recommended for most workflows
Load one population	50--80 GB	80 GB	e.g., all 300k stars
Load full dataset	150--200 GB	80 GB	Only if you have the RAM
HuggingFace streaming	1 GB	0	No local download needed

Recommended workflow

For most users, iterate shard by shard rather than loading everything:

import pandas as pd
import glob

for shard_path in sorted(glob.glob("path/to/shards/*.parquet")):
    df = pd.read_parquet(shard_path)
    # Filter, process, extract features...
    del df

Or use HuggingFace streaming to avoid downloading at all:

from datasets import load_dataset
ds = load_dataset("kshitijd/astro-multimodal-570k", streaming=True)
for row in ds["train"]:
    # process row...
    pass

---

How It Was Built

Pipeline Overview

Built with a custom Python pipeline running on NCSA Delta AI (32 CPUs, 408 GB RAM, NVMe storage). Total runtime: ~20 hours.

Phase 1: Catalog Construction

Three populations were defined from independent parent catalogs:

1. Stars (300k): Queried APOGEE DR17 via VizieR (catalog III/284, SNR > 50), cross-matched against Gaia DR3 via CDS XMatch (1 arcsec radius, 50k-row chunks). Selected the first 300k Gaia-matched stars.

2. Galaxies (200k): Downloaded the PROVABGS seed catalog from the Flatiron Institute (60 HDF5 cells). Selected the first 200k objects.

3. AGN (100k): Queried the SDSS DR14 Quasar Catalog via VizieR (catalog VII/286). Selected the first 100k objects.

Phase 2: Flatiron Cross-Matching (Gaia, TESS, SDSS, DESI)

For each Flatiron-hosted dataset, the pipeline downloaded HEALPix-partitioned HDF5 files, performed positional cross-matching with astropy.coordinates.SkyCoord (3 arcsec radius), accumulated matched data, and flushed to Parquet shards. Each HDF5 file was deleted after processing.

Phase 3: APOGEE Spectra

APOGEE aspcapStar FITS files (73 GB) were pre-staged via Globus. The pipeline read each star's local FITS file and cropped spectra to good detector regions: np.r_[246:3274, 3585:6080, 6344:8335] (7514 pixels).

Phase 4: SDSS Spectra

SDSS specLite files were downloaded from the SDSS Science Archive Server for galaxies and AGN with matching plate/MJD/fiber identifiers.

Phase 5: ZTF Light Curves

ZTF DR24 bulk Parquet files from IRSA. Per-field download, PyArrow read, group by object ID, positional cross-match, extract time series.

Phase 6: Images (concurrent)

Four image sources ran concurrently:

• 2MASS: CDS hips2fits (J, H, K bands)
• GALEX: CDS hips2fits (FUV, NUV bands)
• Legacy Survey: Cutout service (g, r, z bands)
• unWISE: Tile download + astropy Cutout2D (W1, W2 bands)

Phase 7: GALAH Spectra

GALAH DR4 spectra (504 tar files, 123 GB) downloaded from Data Central Australia. Each star's 4-band spectra (blue, green, red, IR) were concatenated.

Phase 8: Finalize

All modality shards merged per population via streaming left joins on object_id. Deduplicated. Modality counts computed.

Cross-Matching

All positional cross-matches used astropy.coordinates.SkyCoord.match_to_catalog_sky() with a 3 arcsecond match radius. Unmatched columns are null.

---

Data Quality Notes

• APOGEE spectra are continuum-normalized. Flux values are typically 0.5--1.2.
• Gaia BP/RP is stored as spectral coefficients (flatiron_gaia_coeff, 110 values), not sampled spectra. See the Gaia documentation for reconstruction into a full spectrum.
• TESS and ZTF light curves have variable lengths depending on sky coverage overlap with the APOGEE/AGN footprints.
• Image cutouts are centered on the catalog position. Some may contain NaN pixels at edges.
• Columns with no data for a given object are null / None.

Known Limitations

1. Galaxy spectral coverage is 20.5%. Most PROVABGS galaxies lack spectroscopic observations. DESI metadata is present for 75% but actual spectra for only 12%.
2. Light curve coverage is partial (TESS ~14%, ZTF ~9% of stars), driven by sky overlap with the APOGEE footprint.
3. unWISE coverage is 7.6% for stars and Legacy Survey is < 1% due to limited footprint overlap.
4. Gaia BP/RP is ~49% for stars, reflecting overlap between APOGEE targets and Flatiron-hosted HEALPix cells.
5. No light curves for galaxies by design -- TESS and ZTF are routed only to stars and AGN.
6. ~340 columns total, most being Gaia and DESI metadata. Core science columns are listed in the Schema section.

---

Citation

If you use this dataset, please cite the underlying surveys:

@article{abdurrouf2022,
  title={The Seventeenth Data Release of the Sloan Digital Sky Surveys: Complete Release of MaNGA, MaStar, and APOGEE-2 Data},
  author={Abdurro'uf and others},
  journal={ApJS},
  volume={259},
  pages={35},
  year={2022}
}

@article{gaia2023,
  title={Gaia Data Release 3: Summary of the content and survey properties},
  author={{Gaia Collaboration}},
  journal={A\&A},
  volume={674},
  pages={A1},
  year={2023}
}

@article{desi2024,
  title={DESI 2024 III: Baryon Acoustic Oscillations from Galaxies and Quasars},
  author={{DESI Collaboration}},
  journal={AJ},
  year={2024}
}

@article{bellm2019,
  title={The Zwicky Transient Facility: System Overview, Performance, and First Results},
  author={Bellm, Eric C. and others},
  journal={PASP},
  volume={131},
  pages={018002},
  year={2019}
}

@article{buder2024,
  title={The GALAH Survey: Data Release 4},
  author={Buder, Sven and others},
  journal={arXiv preprint arXiv:2409.19858},
  year={2024}
}

@article{ricker2015,
  title={Transiting Exoplanet Survey Satellite (TESS)},
  author={Ricker, George R. and others},
  journal={Journal of Astronomical Telescopes, Instruments, and Systems},
  volume={1},
  pages={014003},
  year={2015}
}

License

Released under CC-BY-4.0. The underlying survey data is subject to each survey's individual data use policies.