StaticCarveCarve the locked holdout off matrix.
the assembled rows (rows with no label already removed)
the pipeline's LockedHoldoutFraction
the label column; when it is not in the matrix the carve is a plain seeded sample
classification carves are stratified on the label; regression carves are not
the split seed (normally DEFAULT_SPLIT_SEED)
StaticFromThe split a Training Snapshot's row manifest records: the development rows are the training portion and the locked-holdout rows the holdout, each in matrix order, with every snapshot column. Nothing is re-carved, so the counts and the holdout class counts describe exactly the rows the manifest names.
the snapshot content (matrix row i is manifest line i)
the label column the holdout class counts are read from
classification counts the holdout rows per label; regression stores no class counts
StaticHoldoutHow many rows a fraction holds out of total: at least one when the fraction is positive, never every row (one
row always stays for training), and none when there is at most one row.
StaticMulberry32Mulberry32: a small, fast 32-bit seeded generator returning floats in [0, 1). Local so the carve has no dependency whose algorithm could change under it — the same seed must give the same holdout across releases.
Seeded, stratified locked-holdout carve. Pure; no I/O.