kshitijd/mmu-l2-sdss

L2 model-ready views (release v1)

The mmu-l2-* repositories provide processed, model-ready versions of the matching mmu-norm-* L1 data, while L1 keeps the normalized source measurements. L2 applies documented processing steps for training and evaluation, with the details for reversing each transformation stored in the row or in provenance.json.

Repo	View	Reversal
mmu-l2-tess	per-sector relative flux f/median−1, time from first valid cadence	flux = (rel+1)×flux_median_e_per_s; t = t_rel + t0
mmu-l2-sdss	spectra zero-padded to 4,800 samples with valid mask	drop valid=false samples
mmu-l2-chandra	bins zero-padded to 512 with valid mask	drop padding
mmu-l2-sne	time-sorted sequences, band/survey tokens, ONE scale per object (flux: ÷ median |f|; mag: − median m), representation column	flux rows: v×scale; mag rows: v+scale
mmu-l2-provabgs	16/50/84 posterior percentiles per parameter	samples remain in L1

Each row keeps its assignment from splits/v1 (train, val, test, or unassigned). Images, DESI and VIPERS spectra, and Gaia already have a fixed shape, so their L2 preparation uses the L1 tensors with the training-only robust scaling statistics in mmu-norm-index under l2_stats/v1/ (z = (x − median) / iqr).

In every L2 view, padding is marked valid=false, one amplitude scale is used per object to preserve colors, statistics are fit on the training split only, and gaps are left as gaps.