astronolan/galaxy10-aion-embedded

Galaxy10 AION-1 Benchmark — Test Patch Activations

This dataset preserves only the 796-row test split, with all original test columns unchanged, used to produce the Galaxy Morphology Classification results in the AION-1 paper (Table 2, Section 7.2.2).

Task

Classify galaxy images into 10 morphology classes from Galaxy Zoo DECaLS:

Label	Class Name
0	Disturbed Galaxies
1	Merging Galaxies
2	Round Smooth Galaxies
3	In-between Round Smooth Galaxies
4	Cigar Shaped Smooth Galaxies
5	Barred Spiral Galaxies
6	Unbarred Tight Spiral Galaxies
7	Unbarred Loose Spiral Galaxies
8	Edge-on Galaxies without Bulge
9	Edge-on Galaxies with Bulge

Splits

Split	Count	Description
test	796	10% held-out evaluation set

These are the exact test indices used in the AION-1 paper. The training split is omitted.

Columns

Column	Type	Description
image_rgb	image	256x256x3 uint8 RGB from Galaxy10_DECals.h5 (PNG)
ra	float64	Right ascension (degrees)
dec	float64	Declination (degrees)
Galaxy10_DECals_index	int64	Row index into Galaxy10_DECals.h5
label	int64	Morphology class (0-9)
label_name	string	Human-readable class name
image_bands	binary	96x96 4-band (g,r,i,z) float32 cutout from Legacy Survey DR10

Reconstructing image_bands

The image_bands column stores float32 flux values as a nested list (4 bands x 96 x 96 pixels). To reconstruct:

import numpy as np
cutout = np.array(row["image_bands"], dtype=np.float32)  # (4, 96, 96)
# cutout[0] = g-band, cutout[1] = r-band, cutout[2] = i-band, cutout[3] = z-band

Using with AION-1

To tokenize with the AION codec and compute embeddings:

import torch
from aion.codecs import CodecManager
from aion.modalities import LegacySurveyImage
from aion.model import AION

codec_manager = CodecManager(device="cuda")
model = AION.from_pretrained("polymathic-ai/aion-base").to("cuda").eval()

cutout = np.array(row["image_bands"], dtype=np.float32)  # (4, 96, 96)
image_flux = torch.tensor(cutout).unsqueeze(0).to("cuda")
image = LegacySurveyImage(flux=image_flux, bands=["DES-G", "DES-R", "DES-I", "DES-Z"])
tokens = codec_manager.encode(image)
embeddings = model.encode({"tok_image": tokens["tok_image"]}, num_encoder_tokens=600)
patch_embeddings = embeddings  # (1, 576, 768); keep every spatial token

Model	Accuracy (%)
EfficientNet	80.0
AION-B + MLP Probe	84.0
AION-L + MLP Probe	87.2
AION-XL + MLP Probe	86.5
DINOv2 + MLP Probe	71.4
ZooBot + MLP Probe	89.6
Human (Author N.K.)	90.1
GPT-4.1-mini (zero-shot + judge)	50.8

Data Sources

• Galaxy10_DECals.h5: Galaxy Zoo 10 catalog (Leung and Bovy, 2018; Walmsley et al., 2022)
• 4-band cutouts: Downloaded from ALASKY via CDS HiPS2FITS at 0.262 arcsec/pixel

The underlying image data is from the DESI Legacy Imaging Surveys and Galaxy Zoo, subject to their respective data-use policies.

Final-layer patch activations

The following float32 columns retain every spatial token in row-major order. No image-level pooling or additional L2 normalization is applied.

Column	Shape per image	Spatial grid	Activation
aion_embedding	576 × 768	24 × 24	polymathic-ai/aion-base: unpooled AION.encode() output after encoder block 12, final normalization, context projection, and positional addition. Input is the original 4 × 96 × 96 image_bands, ordered DES-G, DES-R, DES-I, DES-Z.
embeddinggemma2_embedding	256 × 768	16 × 16	google/embeddinggemma-2: image-token entries in last_hidden_state, after text-encoder layer 24, final RMSNorm, and the 512→768 projection. All special and padding tokens are excluded. The default 280-soft-token maximum yields 256 actual image tokens for square images; no task prompt is used.
dinov3vitl_embedding	196 × 1024	14 × 14	facebook/dinov3-vitl16-pretrain-lvd1689m: final-layer, model-normalized patch entries in last_hidden_state after block 24. The CLS token and four register tokens are excluded. The published RGB processor resizes to 224 × 224.

EmbeddingGemma 2 and DINOv3 use the original image_rgb values with their published processors. The model's internal final normalization is retained; no extra vector normalization is applied.

These patch activations support fitting block-sparse featurizers as described in Structuring Sparsity: Block-Sparse Featurizers Capture Visual Concept Manifolds. Each image remains one dataset row; flatten the first two tensor axes across rows when assembling per-patch training matrices. Spatial grids are retained for concept maps. The paper's DINO experiment uses ViT-B; this dataset uses the requested ViT-L checkpoint.

import numpy as np
from datasets import load_dataset

test = load_dataset("astronolan/galaxy10-aion-embedded", split="test", streaming=True)
row = next(iter(test))
aion = np.asarray(row["aion_embedding"], dtype=np.float32)  # (576, 768)
gemma = np.asarray(row["embeddinggemma2_embedding"], dtype=np.float32)  # (256, 768)
dino = np.asarray(row["dinov3vitl_embedding"], dtype=np.float32)  # (196, 1024)

Pinned revisions

Source dataset: astronolan/galaxy10-aion@81afc8d8b83555a51e0e43260a5dae4c4e4db7df.

• polymathic-ai/aion-base@40541618104bab0fa85c8af68daeb867a720bb8c
• google/embeddinggemma-2@914f7f89142e33e77833254d9c9b90c3cef7303b
• facebook/dinov3-vitl16-pretrain-lvd1689m@ea8dc2863c51be0a264bab82070e3e8836b02d51