BrunoBarreto/sdss_dr17_spectra_for_classification

SDSS DR17 Spectral Classification Parquet Benchmark

This fixed supervised benchmark classifies SDSS DR17 optical spectra as GALAXY, QSO, or STAR. Rows contain native FITS arrays, common-grid spectral inputs, identifiers, derived spectral-quality fields, and metadata matched from specObj-dr17.fits.

Files

train.parquet
validation.parquet
test.parquet
dataset_metadata.json

Split	Number of spectra
Train	22,581
Validation	4,839
Test	4,839

The stratified split preserves the original 70%/15%/15% train/validation/test assignment, using seed 42.

Target labels

class_label is the target. class_id mapping:

{
  "GALAXY": 0,
  "QSO": 1,
  "STAR": 2
}

Class	Number of spectra
GALAXY	20,348
QSO	5,803
STAR	6,108

Spectral arrays

native_flux, native_loglam, and native_ivar are the original ragged observed-frame FITS arrays. native_mask is true for non-finite pixels or ivar <= 0. native_and_mask and native_or_mask are SDSS pixel-mask bit vectors. processed_flux and processed_ivar are resampled onto the common grid used by the benchmark, but are not PCA components.

The common grid is stored in dataset_metadata.json: 4,000 points from log10 wavelength 3.6993999 to 3.8018000 (5004.95-6335.78 Angstrom). It is the strict common-support interval, so it does not extrapolate.

To reproduce the classification protocol, fit standardisation and PCA on train.parquet only. The catalog_* fields provide DR17 object IDs, positions, survey/program provenance, class/subclass, redshift diagnostics, warnings, primary flag, and S/N values.

Loading

from datasets import load_dataset
dataset = load_dataset("BrunoBarreto/sdss_dr17_spectral_classification")
train, validation, test = dataset["train"], dataset["validation"], dataset["test"]

For local use:

import pandas as pd
train = pd.read_parquet("train.parquet")
X_train = train["processed_flux"]
y_train = train["class_label"]

Provenance and responsible use

Built from the local SDSS FITS collection and DR17 specObj-dr17.fits catalogue.