BrunoBarreto/sdss_dr12_stars_regression

SDSS DR12 Stellar Spectra 30k/5k/15k Parquet Benchmark

This dataset provides a fixed supervised benchmark split for stellar atmospheric parameter estimation from SDSS DR12 optical spectra.

It contains three Parquet files:

train.parquet
validation.parquet
test.parquet

The split sizes are:

Split	Number of spectra
Train	30,000
Validation	5,000
Test	15,000

Each row corresponds to one SDSS stellar spectrum and includes raw spectral arrays, processed spectral features, source identifiers, basic metadata, and catalog stellar-parameter labels with uncertainties.

Files

train.parquet

Training split used for model fitting.

validation.parquet

Validation split used for model selection and hyperparameter tuning.

test.parquet

Held-out test split used only for final evaluation.

Columns

Column	Description
spec_id	Internal spectrum identifier used to align the split with the downloaded FITS files
plate	SDSS plate identifier
mjd	Modified Julian Date of the observation
fiber	SDSS fiber identifier
ra	Right ascension in degrees
dec	Declination in degrees
snr	Signal-to-noise ratio from the source catalog, when available
rv_adop	Adopted radial velocity from the source catalog
rv_adop_unc	Uncertainty of the adopted radial velocity
flux	Raw SDSS flux array from the FITS spectrum
loglam	Log10 wavelength array corresponding to flux
ivar	Inverse variance array from the FITS spectrum
mask	Boolean mask where invalid or non-positive-inverse-variance pixels are marked
processed_flux	Fixed-length processed spectral feature vector used by the benchmark models
catalog_teff	Adopted catalog effective temperature
catalog_teff_unc	Uncertainty of catalog_teff
catalog_feh	Adopted catalog metallicity [Fe/H]
catalog_feh_unc	Uncertainty of catalog_feh
catalog_logg	Adopted catalog surface gravity
catalog_logg_unc	Uncertainty of catalog_logg

Target Labels

The supervised regression targets are the adopted catalog stellar parameters:

Target	Description	Unit
catalog_teff	Effective temperature	K
catalog_feh	Metallicity relative to solar	dex
catalog_logg	Surface gravity	dex

The corresponding uncertainty columns are:

catalog_teff_unc
catalog_feh_unc
catalog_logg_unc

These are taken from the catalog uncertainty fields:

TEFF_ADOP_UNC
FEH_ADOP_UNC
LOGG_ADOP_UNC

Loading the Dataset

from datasets import load_dataset

dataset = load_dataset("BrunoBarreto/sdss_dr12_stars_regression")

train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

print(train[0].keys())

Example: Training on Processed Flux

import numpy as np

from sklearn.linear_model import RidgeCV
from sklearn.metrics import mean_absolute_error

X_train = np.stack(train["processed_flux"]).astype("float32")
X_test = np.stack(test["processed_flux"]).astype("float32")

y_train = np.array(
    [
        train["catalog_teff"],
        train["catalog_feh"],
        train["catalog_logg"],
    ],
    dtype="float32",
).T

y_test = np.array(
    [
        test["catalog_teff"],
        test["catalog_feh"],
        test["catalog_logg"],
    ],
    dtype="float32",
).T

alphas = np.logspace(-3, 3, 13)

print("Ridge ...")
model = RidgeCV(alphas=alphas)
model.fit(X_train, y_train)

pred = model.predict(X_test)

mae = mean_absolute_error(y_test, pred, multioutput="raw_values")

print(f"MAE Teff:   {mae[0]:.2f} K")
print(f"MAE [Fe/H]: {mae[1]:.3f} dex")
print(f"MAE logg:   {mae[2]:.3f} dex")

The raw arrays can be used to build custom preprocessing pipelines or to feed models that require native observed-frame spectra.

Paper: arxiv.org/abs/2606.13868