YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Remote-sensing native assets
Remote-sensing task bytes are deliberately not stored in Git. The files are licensing-controlled, and evaluator references are host-only. Committed manifests retain their content identities and audience policy.
A QGIS family that carries several AOIs keeps one directory per case, so a case is self-contained:
<family>/<case-id>/inputs/ guest-visible task inputs, uploaded by the task JSON
<family>/<case-id>/expected/ host-only reference, never uploaded to a guest
A family with a single AOI drops the case level and puts inputs/ and expected/ directly under
the family — the case level would only ever hold one entry. qgis-road-centerline-digitization is
the first family shaped this way. The task id still carries its case-<hash> suffix, because that
is the OSWorld example identity every QGIS task is indexed by, not a directory name.
.gitignore excludes every byte below evaluation_examples/assets/; only *.md and
remote_sensing/manifests/ stay tracked. That rule is what keeps a reference out of Git, so it
must hold for any new family without further edits. SNAP still keeps its protected reference under
private/ and has not been migrated to this layout.
The negations that sentence promises are not in
.gitignore. Line 228 is a bare/evaluation_examples/assets/remote_sensing/**with nothing re-included after it, so the tracked*.mdandmanifests/*.jsonsurvive only because they were already tracked when the rule landed. A new manifest or note added under this directory is silently ignored — verify withgit check-ignore -v <path>and add the negation lines before relying on it.
Hydrate all five permanent building-damage cases from the maintainer-held archive:
python scripts/ops/hydrate_sci_assets.py \
--archive-root /mnt/e/WSL-backups/OSWorld-native-migration-assets-20260813
For a verified local ArtifactWorkspace copy, use --cas-root with the directory that contains
objects/sha256. The command copies rather than moves, verifies the archive and every selected
object, refuses conflicting existing files, and writes the protected reference only below the
case's expected/ directory.
To hydrate only one case for a focused check, pass --case-id case-...; the option is repeatable.
The frozen family manifest has SHA-256
2d0f750ce2ab4e2352d7e36875e974d5569f5ba17aded13d08f072c04e5fcd67.
Candidate evidence remains
in the archive but is never hydrated because native outcome scoring does not consume it. The three
controlled inputs for each case are written below qgis-building-damage/<case-id>/inputs/; each
protected reference is written mode 0600 to qgis-building-damage/<case-id>/expected/ and is
never uploaded to a guest.
qgis-road-centerline-digitization
Not in the maintainer archive and not hydratable — the two files were copied by hand from
RSCUA-Bench on 2026-08-20, where they were untracked at the time. Re-copy from
RSCUA-Bench/data/road_centerline_digitization_01/ and check the digests below.
| File | SHA-256 | Bytes |
|---|---|---|
inputs/satellite.tif |
dbcbbc11e682aabdddcf0e9d2c62974e00d49c01f0d5a7ec6127641b869adebb |
1503122 |
expected/reference.json |
1206c9628979e5dfabf3b662bc5a8d943d73846342d5954ca19175900729eb11 |
763 |
The reference digest is also pinned in the task's evaluator.options.reference_sha256, and both
the evaluator and the oracle driver refuse to run when it moves.
Source: SpaceNet3 AOI_2_Vegas img82, cropped to 734x734 px at 0.2726 m/px, EPSG:32611 — roughly
200 m x 200 m of detached housing. The upstream SpaceNet3 licence terms, the crop parameters, and
the digests of the original tiles are not recorded anywhere; that provenance was already
missing in RSCUA-Bench and did not travel with the bytes.
- Downloads last month
- 14