asterisk-labs/rumi-api-fixtures

Rumi API Fixtures

A small, deterministic collection of Earth-observation arrays encoded as Rumi files.

This repository exists to test the Rumi API and its stateless remote range reads through Karu. It is not a training dataset, a scientific benchmark, or a general-purpose imagery archive.

What this dataset tests

The fixtures cover:

• complete and windowed reads;
• band selection and ordering;
• temporal selection;
• batch reads with rumi.read_many;
• different Rumi frame layouts;
• signed and unsigned integer types;
• single-band, multiband, and high-dimensional arrays;
• geospatial and temporal metadata;
• HTTP Range reads from Hugging Face.

Repository structure

data/             Rumi containers
headers/          external binary headers required for remote reads
verify/           local and Hugging Face verification programs
manifest.json     source identity, storage metadata, and expected results
checksums.sha256  SHA-256 checksums for every container and header
SOURCE_DATA.md    source attribution and licensing notes

Each .rumi file has a corresponding external header. The manifest records its original EarthCompress sample, logical shape, data type, frame layout, tile size, metadata, and checksums.

Reading a fixture

from pathlib import Path

import rumi
from huggingface_hub import hf_hub_download

repo = "asterisk-labs/rumi-api-fixtures"
name = "s2-00-tile"

header_path = hf_hub_download(
    repo_id=repo,
    repo_type="dataset",
    filename=f"headers/{name}.header",
)

header = Path(header_path).read_bytes()
image = rumi.read(
    f"hf://datasets/{repo}/data/{name}.rumi",
    header,
    bands=[0, 3],
    window=(0, 0, 256, 256),
)

print(image.shape)

For reproducible tests, replace the default repository revision with a pinned commit or release tag.

Fixtures

Fixture	Source corpus	Purpose
s2-00-tile	Sentinel-2 L1C	Window and band-selective reads
s2-01-tile	Sentinel-2 L1C	Batch reads
s2-02-tile	Sentinel-2 L1C	Batch reads
s2-00-planar	Sentinel-2 L1C	Planar frame layout
s2-00-chunky	Sentinel-2 L1C	Pixel-interleaved frame layout
era5-t2m-00-time	ERA5	Temporal selections and ragged edges
alphaearth-00	AlphaEarth	Signed int8 with 64 bands
emit-00	EMIT L2A	Signed int16 with 285 bands
s1-00	Sentinel-1 GRD	Signed radar values
worldcover-00	ESA WorldCover	Single-band categorical data

The three encodings of s2-00 contain the same logical array. They are intentionally repeated to verify that frame layout changes storage and range behavior without changing decoded results.

Verification

Validate the generated files locally:

python verify/verify_local.py

After publication, validate actual remote range reads:

python verify/verify_huggingface.py --revision main

The remote verifier downloads only the manifest and the small external headers through huggingface_hub. Rumi reads the selected array windows directly from the remote .rumi objects.

Data provenance

The arrays are selected from the EarthCompress benchmark corpora. Their values are preserved; only their storage representation changes when encoded as Rumi files.

manifest.json identifies the exact source corpus and sample for every fixture and preserves the original array checksum. See SOURCE_DATA.md for the required source attributions.

Licensing

The underlying observations remain subject to the terms of their respective data providers. The applicable attribution and redistribution notes are recorded in SOURCE_DATA.md.

Scope

These fixtures are deliberately small and are not statistically representative of their source datasets. Passing these tests establishes API and format compatibility; it does not measure compression performance or scientific fitness.