gaia-eb-teff-models

Gaia Eclipsing Binary Effective Temperature Models

Pre-trained Random Forest models and a soft PCA–KDE ensemble for predicting (T_{\mathrm{eff}}) of Gaia DR3 eclipsing binaries from photometry, extinction, and EB catalogue quantities (no parallax). Training labels are polynomial-corrected GSP-Phot temperatures (hot-star blend above 10 000 K).

Code: gabdeevmaksim/gaia_eb_rf_teff

Production models (July 2026)

Ensemble: ensemble_selector_gaia_teff_production_soft_kde_20260715_140546.pkl (soft weights in a 2D PCA plane; MSC off; EB-params out-of-coverage fallback).

Soft-KDE holdout (common ~28k sample)

Catalogue: 847 486 ML fills; combined with GSP-Phot → 97.9% Teff coverage of (2,169,721) cross-matched EBs. Do not use older “263 K Best-of-Three” claims from earlier Hub versions.

Quick download

pip install huggingface_hub joblib scikit-learn polars
# From the git repo:
python scripts/download_datasets.py --models all

from huggingface_hub import hf_hub_download
import joblib

path = hf_hub_download(
    repo_id="Dedulek/gaia-eb-teff-models",
    filename="rf_gaia_teff_base_ebv_freq_optuna_20260712_195904.pkl",
    repo_type="model",
)
model = joblib.load(path)

Features for the base/flag1 models (order matters): g, bp, rp, bp_rp, g_bp, g_rp, ebv_sandf, frequency. Targets are trained in (\log_{10}(T_{\mathrm{eff}}/\mathrm{K})); invert with 10 ** pred.

Cluster and EB-params models need their sidecar .pkl files (kmeans/scaler or categorical encoder) listed in model_registry.yaml.

Legacy

Older Nov‑2025 models (gaia_teff_corrected_log, 2MASS, etc.) may remain under legacy/ on this Hub for reproducibility of prior experiments. Prefer the July 2026 production quartet + soft-KDE selector for new work.

Citation

Please cite the accompanying paper and this repository when using these models. Gaia data remain subject to ESA archive terms.