split_summary.json

{
  "pilot": "magick_transparency_v1",
  "cache": "cache_named",
  "val_split_ratio": 0.05,
  "method": "deterministic: sha1(f'{dataset_name}:{sample_id_abs_path}')[:8]/2**64; <r->val, <2r->test, else train (matches trainer _record_split)",
  "note": "split is PER-SAMPLE (npz path), not per-object; the same object can appear in different splits.",
  "total": 4380,
  "counts": {
    "train": 3932,
    "val": 232,
    "test": 216
  },
  "counts_by_dataset_name": {
    "test|wsg_named": 216,
    "train|wsg_named": 3932,
    "val|wsg_named": 232
  },
  "unique_objects_per_split": {
    "train": 0,
    "val": 0,
    "test": 0
  },
  "objects_overlap_train_val": 0,
  "objects_overlap_train_test": 0
}