---
pretty_name: "AION-Search Streams, Shells, and Hard Negatives"
license: mit
tags:
  - astronomy
  - stellar-streams
  - shell-galaxies
  - tidal-features
  - semantic-search
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
      - split: test
        path: data/test-*.parquet
dataset_info:
  features:
    - name: image_rgb
      dtype: image
    - name: image_stretched
      dtype: image
    - name: image_bands
      list:
        list:
          list:
            dtype: float32
    - name: ra
      dtype: float64
    - name: dec
      dtype: float64
    - name: label
      dtype: string
  splits:
    - name: train
      num_examples: 72
    - name: test
      num_examples: 30
---

# AION-Search streams, shells, and hard negatives

Catalog with 102 objects: 39 streams, 34
shells, and 29 negatives.

## Splits

The test split is a reproducible stratified random sample with exactly 10
streams, 10 shells, and 10 neither objects. The remaining 72 objects form the
training split (29 streams, 24 shells, 19 neither). Random seed: `20260802`.

## Columns

- `image_rgb`: 256 x 256 Legacy Survey DR10 color JPEG.
- `image_stretched`: the same RGB cutout after the AION-Search per-channel
  arcsinh stretch, grayscale conversion, inversion, and a 0.7-pixel Gaussian blur.
- `image_bands`: Legacy Survey DR10 flux planes as nested float32 lists with
  shape `(4, 256, 256)` and band order `(g, r, i, z)`.
- `ra`, `dec`: ICRS coordinates in decimal degrees.
- `label`: one of `stream`, `shell`, or `neither`.

Every representation uses the same 0.035 degree sky footprint and 256 x 256
pixel grid from CDS HiPS2FITS.

## Curation

The starting catalog contained 114 objects (36 streams, 38 shells, 40 neither).
After a randomized blind review, 12 ambiguous objects were dropped and four
objects were relabeled. The retained catalog is deduplicated at 10 arcsec.
`neither` is an operational hard-negative label from rejected high-ranking shell
semantic-search results; it is not proof of the physical absence of all tidal
structure.

Underlying imagery is from the DESI Legacy Imaging Surveys DR10 and remains
subject to its original data-use policy and acknowledgements.