splits/split_stats.json

{
  "train_count": 10000000,
  "val_count": 500000,
  "test_count": 500000,
  "split_format": "hf_parquet_uint16_target_plus_padded_meta_v1",
  "subset_policy": {
    "train_prefix_count": 10000000,
    "val_prefix_count": 500000,
    "test_prefix_count": 500000
  },
  "rows_per_file": 32768,
  "compression": "zstd"
}