LuciexJune/ome-vip-index

Zarr Streaming Index Dataset

This dataset is a lightweight index for Zarr / OME-Zarr stores hosted on the Hugging Face Hub.

Instead of storing large multidimensional arrays directly in dataset rows, this repository stores one row per sample with metadata and an hf:// path to the corresponding Zarr store. The actual array data remains in a separate data repository and is loaded lazily through the datasets.Zarr feature.

This pattern is useful for large scientific imaging datasets where the raw data is too large to materialize eagerly and should be accessed by slice, region, multiscale level, thumbnail, or training patch.

Dataset Structure

Each row contains a reference to one Zarr or OME-Zarr store.

Example columns:

column	description
sample_id	Stable sample identifier
scan	hf:// path to the Zarr / OME-Zarr store
label	Optional class or condition label
split	Optional split metadata
additional metadata	Optional experiment, subject, tissue, channel, or acquisition metadata

The scan column is intended to be cast to datasets.Zarr.

Recommended Usage

from datasets import Zarr, load_dataset

ds = load_dataset(
    "USERNAME/INDEX_REPO_NAME",
    split="train",
    streaming=True,
)

ds = ds.cast_column("scan", Zarr())

sample = next(iter(ds))
proxy = sample["scan"]

print(type(proxy))
print(proxy.shape)
print(proxy.dtype)