bamfai/galaxy-chirality-catalog

---
license: cc-by-4.0
task_categories:
- image-classification
- tabular-classification
tags:
- astronomy
- galaxies
- chirality
- cosmology
- parity-violation
- equivariant
- deep-learning
size_categories:
- 1M<n<10M
configs:
- config_name: default
data_files:
- split: train
path: catalog_production.parquet
dataset_info:
features:
- name: dr8_id
dtype: int64
- name: p_cw_eq
dtype: float32
- name: p_ccw_eq
dtype: float32
- name: p_ns_eq
dtype: float32
- name: class_eq
dtype: string
- name: confidence_eq
dtype: float32
- name: ra
dtype: float64
- name: dec
dtype: float64
- name: p_cw_raw
dtype: float32
- name: p_ccw_raw
dtype: float32
- name: p_ns_raw
dtype: float32
- name: class_raw
dtype: string
- name: confidence_raw
dtype: float32
- name: image_url
dtype: string
splits:
- name: train
num_examples: 8474531
---

# Galaxy Chirality Catalog — 8.47M Galaxies

A production catalog of **8,474,531 galaxies** from the DESI Legacy Survey DR8, classified by visual chirality (spin direction) into three classes: **clockwise (CW)**, **counter-clockwise (CCW)**, and **not spiral (NOT_SPIRAL)**.

This catalog was produced using test-time equivariant averaging to eliminate all optical and handedness biases from the classifier. All 8 of 8 bias-validation tests pass, confirming that the catalog is free of systematic chirality bias.

## Key Results

| Metric | Value |
|--------|-------|
| Total galaxies | 8,474,531 |
| CW galaxies | 1,592,107 |
| CCW galaxies | 1,609,053 |
| Not-spiral galaxies | 5,273,371 |
| CW / (CW + CCW) | **0.4974** |
| Dipole amplitude | 0.0044 |
| Dipole significance | **0.43σ (null)** |
| Dipole p-value | 0.30 |
| Dipole direction (l, b) | (293.0°, 12.0°) |

The equivariant CW fraction of 0.4974 is consistent with parity symmetry. The dipole search over high-confidence spirals (n = 949,584) yields an amplitude of 0.43σ — fully consistent with isotropy and no preferred cosmic spin axis.

## Column Descriptions

| Column | Type | Description |
|--------|------|-------------|
| `dr8_id` | int64 | Unique galaxy identifier from DESI Legacy Survey DR8 |
| `p_cw_eq` | float32 | Equivariant probability of clockwise spiral |
| `p_ccw_eq` | float32 | Equivariant probability of counter-clockwise spiral |
| `p_ns_eq` | float32 | Equivariant probability of not-spiral |
| `class_eq` | string | Equivariant predicted class (CW / CCW / NOT_SPIRAL) |
| `confidence_eq` | float32 | Equivariant classification confidence (max probability) |
| `ra` | float64 | Right ascension (degrees, J2000) |
| `dec` | float64 | Declination (degrees, J2000) |
| `p_cw_raw` | float32 | Raw (single-pass) probability of clockwise spiral |
| `p_ccw_raw` | float32 | Raw (single-pass) probability of counter-clockwise spiral |
| `p_ns_raw` | float32 | Raw (single-pass) probability of not-spiral |
| `class_raw` | string | Raw predicted class (CW / CCW / NOT_SPIRAL) |
| `confidence_raw` | float32 | Raw classification confidence |
| `image_url` | string | URL to the galaxy cutout image in DESI Legacy Survey |

### Equivariant vs. Raw Predictions

- **Equivariant (`_eq`)**: Test-time averaged over all 8 dihedral-group transformations (4 rotations x 2 reflections). This eliminates any optical handedness bias from the classifier. **Use these columns for science.**
- **Raw (`_raw`)**: Single forward-pass predictions without augmentation. Included for comparison and bias-validation purposes.

## Methodology

1. **Source images**: 8.47M galaxy cutouts from [Smith42/galaxies](https://huggingface.co/datasets/Smith42/galaxies) (DESI Legacy Survey DR8).
2. **Model**: [bamfai/galaxy-chirality-v2](https://huggingface.co/bamfai/galaxy-chirality-v2) — a 3-class ResNet-based classifier trained on the GalaxyMNIST morphology dataset with chirality labels.
3. **Equivariant averaging**: Each galaxy image is transformed under all 8 elements of the dihedral group D4. CW/CCW probabilities are swapped for reflections. The 8 probability vectors are averaged to produce perfectly equivariant predictions.
4. **Bias validation**: 8/8 tests pass, including CW/CCW symmetry, hemisphere balance, magnitude independence, and equivariant consistency checks.
5. **Dipole search**: Spherical harmonic decomposition of the CW excess field over high-confidence spirals, testing for a preferred cosmic axis.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("bamfai/galaxy-chirality-catalog")

# Access the catalog
df = ds["train"].to_pandas()

# High-confidence spirals only
spirals = df[df["class_eq"].isin(["CW", "CCW"]) & (df["confidence_eq"] > 0.7)]
print(f"CW fraction: {(spirals.class_eq == CW).mean():.4f}")
```

## Files

| File | Description |
|------|-------------|
| `catalog_production.parquet` | Full 8.47M galaxy catalog (909 MB) |
| `catalog_c_summary.json` | Summary statistics (counts, fractions, runtime) |
| `dipole_catalog_c.json` | Dipole analysis results (amplitude, direction, significance) |

## Citation

```bibtex
@dataset{golden2026chirality,
author = {Houston Golden},
title = {Galaxy Chirality Catalog: 8.47M Galaxies from DESI Legacy Survey DR8},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/bamfai/galaxy-chirality-catalog},
note = {BigBounce Research, https://bigbounce.hubify.app}
}
```

## License

CC-BY-4.0

## Links

- **Research website**: [bigbounce.hubify.app](https://bigbounce.hubify.app)
- **Model**: [bamfai/galaxy-chirality-v2](https://huggingface.co/bamfai/galaxy-chirality-v2)
- **Source data**: [Smith42/galaxies](https://huggingface.co/datasets/Smith42/galaxies)
- **Author**: Houston Golden ([e-mail])