DISFOR-S1-Companion / docs /PUBLICATION_CHECKLIST.md
grujdin's picture
Stage DISFOR S1 v1.0.0 batch 4
4dac7ae verified
|
Raw History Blame Contribute Delete
11.1 kB

From a completed local candidate to Hugging Face publication

This builder prepares a local release candidate. It does not create a Hugging Face repository, upload files, change repository visibility, assign authorship or choose a licence. The current next action is to finish and inspect the Windows build. Publication metadata and remote validation follow that result.

1. Confirm the completed full build

After the Windows build finishes, retain its console result and BUILD_REPORT.json from the separate work directory. In the release directory, inspect release.json and RELEASE_COMPLETE.json.

  • The scope is full, and full_catalogue_release is true. A successful pilot or collected_examples build is still a subset.
  • Counts agree with the accepted source snapshot: 3,777 locations, 136,755 annual blocks, 5,507,831 recorded observations and 5,240,543 accepted centres.
  • The completion marker exists and the package verification command passes. All changed or missing files must be resolved before using the candidate as a publication source.
  • The public reader opens the completed candidate and reads the intended sample/track, including the repaired 3460 / R160_ASCENDING / 2020 block.
  • Samples 1981 and 1998 remain in the catalogue with zero accepted centres. The data dictionary explains their quality status.
  • The exported numerical arrays, masks and native dates have passed construction validation. The earlier hash-only audit and the ten-cube evidence review are described with their actual scopes.

The work directory is a build record, not a publication directory. Preserve it locally with the accepted source audit. No source extraction or old neural-network experiment needs to be restarted merely to proceed with publication.

2. Settle the publication metadata

The following items remain project-owner information; the builder does not guess them:

Item Concrete information to record
Repository owner and name Exact account or organization namespace and dataset repository URL. Use an owner authorized to publish this companion.
Dataset title Final title and the intended version, initially the proposed 1.0.0.
Authors and contributors Names, order, roles, affiliations and optional ORCIDs, as agreed by the contributors.
Contact Maintainer or contact address for data questions and error reports.
Component licences The chosen terms for the companion's new contributions and code, together with applicable original DISFOR, Sentinel-1 and Copernicus DEM notices.
Citation A citation record for this companion and the required upstream attribution. The upstream DISFOR DOI is not the new companion's DOI.
Description Intended uses, construction, spatial/temporal coverage, masks, limitations and distinction between dataset integrity and model performance.

Preserve the upstream revision and metadata hashes. The source DISFOR card declares EUPL-1.2; that is an upstream declaration, not a completed licensing decision for every component of this companion. The final release should contain the selected licence text(s) and component notices. This checklist does not choose or reinterpret those terms.

Use the full-source figures to describe sampled locations, not wall-to-wall European imagery. Retain the distinction between undisturbed forest, physiological health, disturbance type, revegetation and annotation gaps. State that labels support the reference pixel and original intervals. The earlier controlled experiments remain separate research evidence; publication of the data does not require claiming successful universal disturbance detection.

3. Finalize metadata without breaking the checksum record

Do not edit files directly inside a completed, sealed release candidate. The generated README, documentation, metadata and other listed files are included in checksums.sha256. The completion record also hashes release.json and the checksum list. A README edit is therefore a real release change, even if no radar arrays change.

Record the intended author/licence/card changes separately. After the full-build result and those choices are available, the metadata-finalization step must produce a new, reviewable candidate with its changed metadata, regenerated file checksums and matching completion record. This package version does not provide a metadata-reseal command. Rerunning an already completed build verifies its existing contents; it is not a supported way to bless manual edits.

Use the finalized and verified candidate for publication. Keep the previous technical candidate and its build record identifiable. Do not manually remove the completion marker or rewrite hashes merely to silence a failed verification.

4. Create or confirm private Hugging Face staging

Once the exact repository owner/name is known:

  1. Sign into the intended Hugging Face account.
  2. Create a dataset repository under that owner, or open the already created dataset repository.
  3. Verify that the repository is private before staging the candidate.
  4. Complete the selected licence and dataset-card metadata through the finalized local files.
  5. Confirm the account's current storage allowance against the measured release size and file layout. The audit's bytes-read counter is not the upload size.

If a repository already exists, inspect its actual privacy setting. A creation-time private option is not a substitute for checking an existing repository's visibility. Use a separate private staging repository or explicitly managed staging branch for pilot tests, and keep the eventual release file set clean and identifiable.

No repository or visibility action is performed by this package. The owner-specific upload instructions can be filled in after the repository URL and finalized candidate are available.

5. Upload with the current Hugging Face interface

Use a separate Hugging Face tool environment or a standalone CLI; do not upgrade the frozen neural-network environment to obtain upload tools. Authenticate locally using the tool's login mechanism. Tokens do not belong in dataset files, run reports or pasted support messages.

The supported folder-upload routes are hf upload and HfApi.upload_folder, with repository type dataset. The v2 migration removed hf upload-large-folder and HfApi.upload_large_folder; earlier instructions naming those interfaces should not be reused. See the official CLI guide, folder-upload guide and v2 migration guide, checked on 10 November 2026.

Stage a completed pilot first if remote loading has not yet been tested. Upload the finalized full candidate once pilot validation passes. Upload the contents of the release directory, preserving its relative paths. Do not upload the builder work directory, local source extract, credentials, checkpoints or original S2 image archives as if they were part of this companion.

Retain the resulting repository commit identifier. A resumed upload must continue to refer to the same finalized local content; do not alter its payloads or indices during transfer. Repository contents and stored checksums must describe the same candidate.

6. Validate the remote copy

After staging, download the specific resulting commit into a new local validation directory, separate from the source candidate. Use a materialized local download with its relative file layout preserved.

  • Compare the downloaded release's metadata and checksums with the finalized local candidate.
  • Run the package's release verification on the downloaded copy. For the final public snapshot, verification must cover the complete published file set, not only a successful download of the README.
  • Open downloaded payloads with the included reader and check a normal trajectory, a rejected-centre example and the repaired block. Confirm that TAR paths and byte offsets work from this new directory.
  • Check the dataset card and viewer/loading configurations. locations should be the default small table. The S1 NPZ cubes use the companion reader; automatic image preview or generic WebDataset support is not promised.
  • Confirm original and project split files are separate, with the seven original unassigned IDs preserved. Hugging Face names such as catalogue and observations are loading groupings, not automatic application of the scientific splits.
  • Keep original S2 imagery external and join by sample_id. Confirm that examples do not invent one-to-one sensor dates or new ground truth.

Hugging Face's data-file configuration documentation explains how declared configurations map to files. The science-specific partition interpretation is documented in this package's DATA_DICTIONARY.md.

If remote validation exposes a problem, correct it through a new finalized candidate and repeat the checks relevant to that problem. Record the replacement revision rather than silently treating the first upload as the validated release.

7. Publish and record an immutable scientific version

After the finalized metadata and remote copy pass validation:

  1. Make the intended dataset repository public.
  2. Record the exact validated commit hash and create the scientific release tag for that revision, for example v1.0.0 if that remains the agreed first version.
  3. Generate the companion's DOI from the dataset settings for the intended validated revision, with the agreed author metadata. Hugging Face provides DOI creation and version-specific DOI updates; consult its DOI documentation and the settings shown at that time.
  4. Store the repository URL, version tag, commit, DOI, completion-record hash and publication date in the project's release records.

Treat a published scientific version tag as fixed. A later data, annotation, mask or preprocessing correction gets a new described version. The DOI for original DISFOR remains an upstream citation alongside the companion's own citation.

If DOI insertion or another later documentation change modifies sealed files, it is another metadata revision and requires an updated checksum/completion record. The external release record can store the newly assigned DOI without changing the already verified payload snapshot. Do not make an unrecorded manual edit to a version being cited.

What to bring back after the Windows build

The immediate follow-up consists of the final console lines and BUILD_REPORT.json, followed by the exact Hugging Face repository URL and the author/licence/contact information when available. Those inputs are used to finalize the publication metadata and generate owner-specific staging commands. The construction work can proceed now without waiting for those publication details.