---
license: cc-by-4.0
task_categories:
  - tabular-regression
  - feature-extraction
language:
  - en
tags:
  - astronomy
  - desi
  - provabgs
  - galaxies
  - multimodal
  - legacysurvey
  - galaxy-parameters
size_categories:
  - 100K<n<1M
---

# PROVABGS desi_legacy_fusion Dataset

## Summary

**109,991 BGS galaxies** with multimodal data from DESI (spectra + photometry) ×
Legacy Survey (imaging + photometry), matched within 1 arcsec.

| Split | Samples |
|-------|---------|
| train | 87,992 |
| validation | 10,999 |
| test | 11,000 |
| **total** | **109,991** |

## Labels (Galaxy Parameters from PROVABGS SED fitting)

| Column | Description | Units |
|--------|-------------|-------|
| `z_hp` | Redshift | — |
| `log_mstar` | Stellar mass | log(M☉) |
| `tage_mw` | Mass-weighted age | Gyr |
| `log_z_mw` | Metallicity log₁₀(Z_MW) | log(Z) |
| `log_ssfr` | Specific SFR log₁₀(SFR/M★) | log(yr⁻¹) |

## Modalities

| Column | Shape | Description |
|--------|-------|-------------|
| `image_pixels_raw` | (102400,) | LS image (4, 160, 160), des-g/r/i/z, flat float32, nanomaggies |
| `spectrum_flux_raw` | (7781,) | DESI-BGS spectrum, float32 |
| `spectrum_ivar` | (7781,) | Inverse variance, float32 |
| `spectrum_mask` | (7781,) | Bad pixel mask, bool |
| `ls_flux_g/r/z/w1/w2` | scalar | Legacy Survey photometry, float32 |
| `desi_flux_g/r/z/w1/w2` | scalar | DESI photometry, float32 |

Wavelength grid in `wavelength_grid.json`, image channel layout in `image_shape.json`.

## Quick Start

```python
from datasets import load_dataset
import numpy as np

BASE = "/mnt/si0009256k6u/ckdata/aiready/provabgs/hf_dataset"
ds = load_dataset("parquet", data_dir=BASE, streaming=True)

for sample in ds["train"].with_format(type="numpy").take(10):
    img  = sample["image_pixels_raw"].reshape(4, 160, 160)  # float32
    spec = sample["spectrum_flux_raw"]                       # (7781,) float32
    z    = sample["z_hp"]
    logM = sample["log_mstar"]
    age  = sample["tage_mw"]
    ssf  = sample["log_ssfr"]
```

## Notes

- `row_group_size=100` for efficient streaming
- List columns stored as float32 (not float64)
- Normalization deferred to training pipeline
- `log_ssfr = log10(AVG_SFR) - log_mstar`
- `log_z_mw = log10(Z_MW)`