Processed PLAsTiCC Data

The raw light curves have variable lengths and irregular cadence across six LSST bands. Instead of binning into a fixed grid (which destroys temporal structure), we keep every observation and represent it as a feature vector that a recurrent model can consume directly. For each object we: 1. Filter out zero-padded placeholder observations 2. Sort all observations chronologically (across all bands) 3. Compute the time elapsed since the previous observation (delta-t) 4. Normalize flux per object 5. One-hot encode the photometric band 6. Calculate the inverse frequence of the classes This produces a variable-length sequence of 9-dimensional feature vectors per object: `(delta_t, flux, flux_err, u, g, r, i, z, Y)`. The original dataset has been split with an 80/20 statified split to maintain the ratios of the classes between the train and test files. The resulting files are: train.npz - contains the keys: `sequences`, `labels`, `object_ids`, and `class_weights` test.npx - contains the keys: `sequences`, `labels`, `object_ids`