roth1414/galaxy-vit-gz-desi-dirichlet-predictions

Galaxy-ViT — GZ DESI Dirichlet predictions

Per-galaxy Dirichlet-Multinomial concentration parameters (α) for the 10-question / 34-answer Galaxy Zoo DESI decision tree, predicted by a Zoobot ConvNeXt-nano encoder finetuned with a Dirichlet-Multinomial head on the DR8 subset of the mwalmsley/gz_desi_wds labeled split.

Dataset summary

Rows	61,440
Columns	36 (key, dr8_id, alpha_0 … alpha_33)
Format	Apache Parquet
File size	~16 MB
Source images	DECaLS DR8 cutouts at 224×224
Inference time	128 s on RTX 5070 Ti (479 galaxies/s)
SHA-256	recorded in the companion .meta.json sidecar

Each row carries the model's per-answer Dirichlet concentration. The posterior mean for answer i of question q is α_i / sum(α_q); the marginal posterior on the answer probability is Beta(α_i, sum(α_q) − α_i) (used by the analytic credible-interval module in galaxy_vit.inference.posterior).

Live demo

These predictions back a 7-tab interactive demo at https://roth1414-galaxy-vit-demo.hf.space: per-question posterior bars, decision-tree Sankey view, 2-D / 3-D UMAP, RA/Dec sky scatter with an Aladin Lite DR10 embed, cosine-kNN similarity search, and a training-progress animation. The v2 changelog is at CHANGELOG.md; the 90-second walkthrough script is in docs/loom_shotlist.md.

Canonical column order (matches the head's 34-D layout)

Index	Question	Answer
0	smooth-or-featured	smooth
1	smooth-or-featured	featured-or-disk
2	smooth-or-featured	artifact
3	disk-edge-on	yes
4	disk-edge-on	no
5	has-spiral-arms	yes
6	has-spiral-arms	no
7	bar	strong
8	bar	weak
9	bar	no
10	bulge-size	dominant
11	bulge-size	large
12	bulge-size	moderate
13	bulge-size	small
14	bulge-size	none
15	how-rounded	round
16	how-rounded	in-between
17	how-rounded	cigar-shaped
18	edge-on-bulge	boxy
19	edge-on-bulge	none
20	edge-on-bulge	rounded
21	spiral-winding	tight
22	spiral-winding	medium
23	spiral-winding	loose
24-29	spiral-arm-count	1, 2, 3, 4, more-than-4, can't-tell
30	merging	none
31	merging	minor-disturbance
32	merging	major-disturbance
33	merging	merger

This ordering is the galaxy_vit.data.schema.question_index_groups() output, derived from upstream galaxy-datasets and verified against the legacy hardcoded constants at every checkpoint load (T3.1).

Intended use

Research-grade. The model recovers known morphology relations on the labeled validation subset:

• Macro top-1 (T2.3 baseline, cross-entropy): 0.8445 on the DR8 test split, comparable to Walmsley+23's published per-question numbers.
• Macro vote MAE (T3.6 Dirichlet): 0.0883 on the DR8 test split.
• Macro coverage at 95% CI (T3.6 Dirichlet, T-calibrated): 0.93.
• Science direction agreement (T5.2): Spearman bar-vs-bulge correlation matches volunteer-observed sign (model ρ = +0.198, volunteer ρ = +0.146, both p < 10⁻²²).

Not appropriate for downstream scientific claims without re-validation against the volunteer-vote ground truth on your sample of interest.

How to use

from huggingface_hub import hf_hub_download
import pandas as pd

path = hf_hub_download(
    repo_id="roth1414/galaxy-vit-gz-desi-dirichlet-predictions",
    filename="gz_desi_dirichlet_v1.parquet",
    repo_type="dataset",
)
df = pd.read_parquet(path)

# Posterior-mean smooth-or-featured probability (3 answers, indices 0..2):
import numpy as np
sof = df[[f"alpha_{i}" for i in range(3)]].to_numpy()
sof_mean = sof / sof.sum(axis=1, keepdims=True)
df["p_smooth"]           = sof_mean[:, 0]
df["p_featured_or_disk"] = sof_mean[:, 1]
df["p_artifact"]         = sof_mean[:, 2]

For analytic 95% credible intervals, see galaxy_vit.inference.posterior.credible_interval in the source repo.

Training data

The model was finetuned on the DR8 split of mwalmsley/gz_desi_wds under the same canonical question / answer schema published by the galaxy-datasets package. Inference predictions in this release cover the same labeled subset (DR8 train + val + test); the unlabeled 8.67M DESI catalog is not included.

Source code

Repository: github.com/roth1414/galaxy-vit (see commit hashes in the companion .meta.json for the exact code state that produced this parquet).

Citation

@misc{galaxy_vit_dirichlet_v1,
  author = {{Galaxy-ViT contributors}},
  title  = {Galaxy-ViT GZ DESI Dirichlet predictions, v1},
  year   = {2026},
  url    = {https://huggingface.co/datasets/roth1414/galaxy-vit-gz-desi-dirichlet-predictions},
  doi    = {<ZENODO_DOI>},
}

Builds on the Galaxy Zoo DESI volunteer effort (Walmsley et al. 2023) and the Zoobot Foundation-Model architecture (Walmsley et al. 2022).

License

Predictions: CC-BY-4.0. Underlying image data and volunteer votes retain their original licenses (Galaxy Zoo DESI / DECaLS imagery CC-BY-4.0; vote data Walmsley+23 / Zenodo 8331338).