Reference / Glossary

Glossary

This page defines the terms that these docs use, each in one meaning.

calibration factor

The factor c that scales the predicted scatter of one head and parameter, fit after training so that the z-scores of the validation stars have unit spread. See Calibration factors.

checkpoint

The directory a training run writes: the best and final weights, the calibration factors, the configuration, the label normalization and the data split. See File formats.

encoder

The part of a multi-head model that maps one survey’s spectrum to a vector of fixed length (latent_dim). Each survey has its own. See The model.

flag

A quality flag of one label, 0 when the label source raises no warning. A label with a nonzero flag is masked. data.max_flag_bits also masks every label of a star with too many flagged parameters.

has_data

A boolean per star and survey: whether the star has a spectrum from that survey in the batch. The model skips a survey that a star lacks.

head

The part of a multi-head model that predicts the labels of one label source from the trunk’s output. See The model.

ivar

The inverse variance of a flux measurement, 1/σ². A pixel with ivar = 0 carries no information.

label

One stellar parameter or abundance of one star, as measured by a label source, with its uncertainty and flag. The parameters are set by the catalogue, for example teff, logg, fe_h and mg_fe.

label source

A high-resolution survey whose labels the model learns to predict: APOGEE or GALAH. Each label source has its own head.

mask

A channel that marks valid entries with 1 and invalid ones with 0: valid pixels in a spectrum (ivar > 0), valid labels in a label array. Masked entries can hold any value and are never used.

missing vector

The learned input of a head’s log-scatter that stands in for a missing label error. See Uncertainties.

normalized units

The units the model predicts in: each parameter minus its training median, divided by its training interquartile range, with Teff in log10. See Preprocessing.

predicted scatter

The model’s estimate s of the spread of the measured labels around its prediction, label noise included: s = sqrt(exp(2 ln s) + s₀²), where ln s is the model output and s₀ the scatter floor.

prediction

The model’s value μ of a label, in normalized units or, after the inverse normalization, in physical units.

scatter floor

The smallest predicted scatter, s₀ (training.scatter_floor, in normalized units).

spectrum

One observation of one star by one survey: its flux and ivar on the survey’s wavelength grid.

split

The division of the usable stars into training, validation and test sets, stored in the checkpoint as row indices. See The data split.

super-catalogue

The HDF5 file with every star’s spectra and labels, one row per star. See What is in the super-catalogue.

survey

A spectroscopic survey whose spectra are model inputs: DESI, BOSS, LAMOST LRS or LAMOST MRS.

trunk

The part of a multi-head model that combines the encoders’ vectors and passes its output to every head.

uncertainty

The spread that DOROTHY reports with a prediction: c s, the calibration factor times the predicted scatter, or s for a checkpoint without calibration factors. See Calibration factors.

z-score

The residual of a prediction divided by its predicted scatter: z = (μ − y) / s. See Z-scores.