Corustella/AstroCLIP-lightweight

AstroCLIP Lightweight 30K

This repository contains the deterministic 30,000-object subset used by the lightweight AstroCLIP backbone experiments in AstroCLIP-lightweight.

Dataset structure

The dataset contains two non-overlapping splits:

Split	Rows
train	24,000
test	6,000

Each row contains exactly three fields:

• image: a float32 array with shape (152, 152, 3);
• spectrum: a float32 array with shape (7781, 1);
• targetid: an int64 astronomical object identifier.

The dataset intentionally does not include a redshift field because this is the exact input dataset used by the reported lightweight-model experiments.

Construction

The source is the MIT-licensed mhsotoudeh/astroclip dataset at source revision 9f36fcb2acc304c7b4fd73d452673d82ba9bf1fb.

Rows were selected deterministically by hashing seed:targetid with BLAKE2b (8-byte digest), retaining the lowest scores. The train split uses seed 42 and the test split uses seed 43. Source train and test partitions were kept separate, and the resulting targetid sets do not overlap.

Usage

from datasets import load_dataset

dataset = load_dataset("Corustella/AstroCLIP-lightweight")
print(dataset)

Intended use and limitations

This subset is intended for reproducible training, retrieval evaluation, and controlled efficiency comparisons of lightweight AstroCLIP variants. It is a deterministic experimental subset, not a statistically representative sample of all astronomical surveys or galaxy populations. Users requiring redshift labels should obtain them from the appropriately licensed upstream source and join them by targetid.

Attribution

Please cite the upstream AstroCLIP work and dataset when using this subset:

• AstroCLIP: Cross-Modal Pre-Training for Astronomical Foundation Models
• mhsotoudeh/astroclip

The upstream dataset is marked as MIT licensed. This derived subset is published under the same license and retains source attribution.