asterisk-labs/taco-api-fixtures

TACO API Fixtures

A deterministic collection of 50 small TACO datasets whose payloads are all Rumi files. The fixtures exercise TACO contracts, metadata hierarchies, FOLDER and ZIP containers, ZIP partitioning, TACOCAT consolidation, generated locations, and local or remote range reads.

They are API fixtures, not training data or a scientific benchmark. Spatial and temporal metadata generated here is intentionally synthetic.

Matrix

The repository combines ten logical contracts with five physical topologies:

Topology	Purpose
folder	One mutable-directory representation
single-zip	One immutable cloud-optimized ZIP
by-size	Six one-sample ZIP partitions plus TACOCAT
by-split	Train, validation, and test ZIP partitions plus TACOCAT
manual-catalog	Three independently written ZIP partitions consolidated into TACOCAT

The ten contracts cover a null structure, fixed and optional assets, fixed and variable sequences, nested folders, point metadata, STAC, ISTAC, nullable coordinates, folder-level STAC, temporal-only metadata, high-dimensional Rumi arrays, and equivalent Rumi frame layouts.

Every topology is an independently openable dataset entrypoint. ZIP files referenced by a TACOCAT are partitions of that dataset and are not counted as additional matrix entries.

Repository structure

data/<case>/<topology>/     generated TACO datasets
generate/generate.py        deterministic matrix generator
verify/verify_local.py      structural and Rumi decode checks
verify/verify_http.py       loopback HTTP Range checks
verify/verify_huggingface.py remote HTTP Range checks
manifest.json               expected cases, entrypoints, assets, and metadata
checksums.sha256            generated-file checksums
source-lock.json            pinned Rumi fixture source
SOURCE_DATA.md              provenance and synthetic-metadata notice

Every object below a dataset's DATA/ tree is a Rumi container. Structured contracts use .rumi names. The null-structure case necessarily uses TACO's extensionless DATA/<sample-index> path. External Rumi headers are stored in Parquet as rumi:header. TACO's own COLLECTION.json and METADATA/*.parquet control files are present as required by the format.

Neither taco:location nor cozip:location is stored in Parquet. Readers construct locations from the physical payload and its container.

Reproduce and verify

From a checkout next to rumi-api-fixtures:

uv sync
uv run python generate/generate.py --clean
uv run python verify/verify_local.py
uv run python verify/verify_http.py

The workspace development configuration also expects the current TACO and Rumi checkouts at ../benchmark/taco/python and ../rumi/bindings/python. Those local sources ensure the fixtures exercise the unreleased reader changes that own taco:location. Once matching releases are available, consumers can resolve the declared package ranges without the workspace source overrides.

The same commands without uv work in an already configured development environment:

python generate/generate.py --clean
python verify/verify_local.py
python verify/verify_http.py

After publication:

python verify/verify_huggingface.py --revision main

The default remote verifier opens all 40 ZIP-backed entrypoints and decodes every Rumi through an HTTP subfile location. Pass --full-api to additionally repeat the schema, raw-level, wide-layout, and location=False checks that the local verifier already performs.

The generator seed, source revision, and source manifest checksum are pinned. Regeneration preserves the logical datasets and synthetic metadata. ZIP byte checksums may change when ZIP timestamps change, so published revisions should be treated as immutable fixture targets.