Datasets:

NoeFlandre
/

osm-polygon-wikidata-only

@@ -1,40 +1,79 @@
-# osm-polygon-wikidata-only
-![OSM Polygon Wikidata dataset overview](assets/dataset_hero.png)
-Extract polygonal OpenStreetMap features from Geofabrik `.osm.pbf` extracts,
-enrich them with Wikidata, Wikipedia, and Wikivoyage, and publish clean,
-multi-table Hugging Face datasets. The default V1 contract is Wikidata-first;
-the explicit V2 contract also keeps polygons discovered through valid
-multilingual `wikipedia=*` tags, including polygons without a Wikidata QID.
-* **GitHub**: <https://github.com/NoeFlandre/osm-polygon-wikidata-only>
-* **Hugging Face dataset**: <https://huggingface.co/datasets/NoeFlandre/osm-polygon-wikidata-only>
-* **V2 Hugging Face dataset**: <https://huggingface.co/datasets/NoeFlandre/osm-polygon-wikidata-and-wikipedia>
-* **Documentation**: <https://noeflandre.github.io/osm-polygon-wikidata-only/>
-* **Maintainer**: Noé Flandre
-## Dataset versions
-The Hugging Face dataset linked above is **V1**. It is produced by the
-Wikidata-first V1 contract in the [`v1.0.0` GitHub code](https://github.com/NoeFlandre/osm-polygon-wikidata-only/tree/v1.0.0):
-the default V1 workflow selects polygons with a valid `wikidata=*` reference.
-The separate [V2 dataset](https://huggingface.co/datasets/NoeFlandre/osm-polygon-wikidata-and-wikipedia)
-uses the V2 workflow to add polygons discovered through multilingual
-`wikipedia=*` tags without changing V1.
 ## Trackio snapshot
-[View the dataset presentation](https://noeflandre.github.io/osm-polygon-wikidata-only/presentations/dataset.html)
-for a visual overview of the published tables and geographic coverage.
-[View the codebase presentation](https://noeflandre.github.io/osm-polygon-wikidata-only/presentations/codebase.html)
-for a visual overview of the architecture and implementation.
-The finished dataset is recorded in the single public Trackio run
-[`final-dataset-snapshot`](https://huggingface.co/spaces/NoeFlandre/osm-polygon-wikidata-only-trackio).
-It contains exactly three plots: a text-coverage funnel, the top ten Wikipedia
-languages plus `Other languages`, and dataset composition on a logarithmic scale.
-The run is a static snapshot with no pipeline timeline.
 | Snapshot metric | Value |
 | --- | ---: |
@@ -50,676 +89,457 @@ The run is a static snapshot with no pipeline timeline.
 | Polygon/link-table storage | 9.9 GB |
 | Total Parquet storage | 19.2 GB |
-The funnel's language thresholds use the canonical Wikipedia polygon fields.
-Wikivoyage is included in the combined document, word, and dataset-composition totals.
-Documentation: [architecture](docs/architecture.md) ·
-[supported Python API](docs/api.md) · [development](docs/development.md) ·
-[contributing](CONTRIBUTING.md) · [security](SECURITY.md)
----
-## What this project does
-1. Reads Geofabrik `.osm.pbf` files (country / region extracts).
-2. Keeps only the polygonal elements:
-   * **Closed ways** carrying a non-empty `wikidata=*` tag.
-   * **Multipolygon relations** carrying a non-empty `wikidata=*` tag.
-3. Computes geometry metadata per polygon (centroid via
-   equirectangular projection, area in m² and km², bbox, area bucket,
-   primary OSM tag).
-4. Looks up the polygons' Wikidata QIDs (entity, sitelinks,
-   descriptions) and then fetches each linked Wikipedia article
-   (lead text, full plain text, page/revision ID, license,
-   attribution).
-5. Publishes the canonical region tables on the Hugging Face Hub:
-   * `polygons/<stem>.parquet` — one row per polygon.
-   * `wikipedia/documents/<stem>.parquet` — one row per unique Wikipedia article revision.
-   * `polygon_articles/<stem>.parquet` — unified many-to-many
-     polygon↔document links for Wikipedia and Wikivoyage.
-   * `manifests/processed_pbfs.json` — aggregate stats per source PBF.
-6. Adds text and fact tables without reprocessing completed PBFs:
-   * `wikipedia/sections/<stem>.parquet` — section-level Wikipedia text.
-   * `wikivoyage/documents/<stem>.parquet` — full Wikivoyage documents.
-   * `wikivoyage/sections/<stem>.parquet` — section-level Wikivoyage text.
-   * `wikidata/facts/<stem>.parquet` — structured claims for polygon entities.
-7. Publishes one frozen Trackio run with static dataset metrics and three plots.
-V2 is an explicit, isolated workflow. It starts from finalized V1 artifacts,
-scans each source PBF for direct Wikipedia tags, reuses matching V1 documents,
-fetches only missing direct pages, and writes under `processed_v2/`. It uses a
-unified `polygon_document_links/` table with `link_sources` provenance and
-publishes to the separate V2 dataset above. V2 also writes
-`wikipedia/sections/<stem>.parquet` with the exact V1 section schema, reusing
-existing V1 rows and fetching only missing Wikipedia revisions. V1 files and
-the default workflow are never rewritten by a V2 run. V2 keeps a resumable
-reuse index in `cache/v2/v1-index/`; completed Parquet row groups are reused
-after an interruption or later restart. Extraction batches and completed direct
-Wikipedia results and section parses are checkpointed under
-`cache/v2/checkpoints/` and removed only after the region reaches its final
-local commit. A restart rescans at most the source PBF because libosmium does
-not expose a portable byte offset, while already saved extraction and fetch
-work is reused and never published until final reconciliation.
-The generated dataset is published on Hugging Face. Local processing inputs,
-intermediate files, and caches use a configurable data root outside the source
-checkout.
----
-## Repository layout
-```
-.
-├── src/osm_polygon_wikidata_only/
-│   ├── __init__.py
-│   ├── augmentation/    # Wikipedia/Wikivoyage sections and Wikidata facts
-│   ├── cli/             # Argument parsing and command adapters
-│   ├── cli/_sync/       # Fail-closed sync publication/retirement helpers
-│   ├── config/          # DataRoot paths and runtime Settings
-│   ├── domain/          # Pure models, schemas, geometry, and identifiers
-│   ├── enrichment/wikidata/ # Wikidata models, parsing, cache, and transport
-│   ├── enrichment/wikipedia/ # Wikipedia models, parsing, cache, and transport
-│   ├── hf/_dataset_stats/ # Dataset-card statistics internals
-│   ├── hf/_geographic/    # Deterministic H3 visualizations
-│   ├── hf/_publication/   # Publication artifact models and validation
-│   ├── hf/_uploader/      # Hub authorization and upload operations
-│   ├── io/              # PBF, Parquet, manifests, cache, and atomic I/O
-│   ├── pipeline/_link_migration/ # Conversion and journaled migration helpers
-│   ├── pipeline/_wikidata_recovery/ # Audited, resumable repair internals
-│   ├── v2/               # Isolated Wikipedia-tag dataset contract and runner
-│   └── utils/           # JSON, logging, retries, time, and scheduling
-├── tests/               # Unit, integration, contract, and golden tests
-��── pyproject.toml       # Build, dev dependencies, Ruff/ty/pytest config
-└── README.md
-```
-Each top-level sub-package has its own `__init__.py` and a tightly
-focused public API. Cross-package imports go through dotted paths.
----
-## Installation
-Requires Python 3.12+ and [`uv`](https://docs.astral.sh/uv/).
-```bash
-git clone https://github.com/NoeFlandre/osm-polygon-wikidata-only.git
-cd osm-polygon-wikidata-only
-uv sync
-```
-This installs all runtime and development dependencies into a managed
-`.venv`:
-| Runtime | Purpose |
-|---|---|
-| `osmium` | Streaming OSM PBF parser |
-| `datasets` | Hugging Face dataset utilities |
-| `huggingface-hub` | HF Hub client |
-| `pyarrow` | Parquet serialization |
-| `typer`, `rich`, `tqdm` | Read-only operator audit CLI, structured reports, and interactive progress |
-| Dev | Purpose |
-|---|---|
-| `pytest`, `pytest-cov` | Tests |
-| `ruff` | Lint + format |
-| `ty` | Type-check production code and maintained scripts |
-| `pre-commit` | Fast commit-time quality checks |
-| [`just`](https://just.systems/) | Shared local and CI command runner |
----
-## Local data root
-All PBF inputs, intermediate outputs, Hugging Face caches, and local
-parquet/manifest files live under one operator-selected **data root**.
-Resolution order:
-1. `--data-root <path>` CLI flag.
-2. `OSM_POLYGON_DATA_ROOT` environment variable.
-A data root is required, must already exist, and is rejected when it points
-inside the source checkout. There is no silent fallback.
-Default sub-directories under the data root:
-| Sub-directory | Purpose |
-|---|---|
-| `raw/` | Geofabrik `.osm.pbf` files (input) |
-| `processed/polygons/` | Written `polygons/<stem>.parquet` files |
-| `processed/articles/` | Local legacy staging files, retained only until their verified canonical publication succeeds |
-| `processed/wikipedia/documents/` | Canonical Wikipedia documents |
-| `processed/polygon_articles/` | Written `polygon_articles/<stem>.parquet` files |
-| `processed/manifests/` | `processed_pbfs.json` aggregate manifest |
-| `logs/` | Reserved for pipeline logs |
-| `hf_cache/` | Hugging Face client-side cache |
-| `cache/wikidata/`, `cache/wikipedia/` | Per-call JSON cache |
-| `cache/` | Shared cache root |
-Set the data root for a session:
-```bash
-export OSM_POLYGON_DATA_ROOT=/path/to/osm-polygon-data
-```
----
-## Usage
-After `uv sync`, the processing and augmentation commands are:
-```bash
-uv run osm-polygon-wikidata-only sync-dir <dir> [--options]
-uv run osm-polygon-wikidata-only process-pbf <input.pbf> [--options]
-uv run osm-polygon-wikidata-only process-dir  <dir>     [--options]
-uv run osm-polygon-wikidata-only augment-region <stem>  [--options]
-uv run osm-polygon-wikidata-only augment-dir             [--options]
-uv run osm-polygon-wikidata-only-trackio                 [--data-root <path>]
-```
-### Common options
-| Flag | Purpose |
-|---|---|
-| `--data-root <path>` | Override the resolved external data root |
-| `--repo-id <org/name>` | Target Hugging Face repo (default `NoeFlandre/osm-polygon-wikidata-only`) |
-| `--user-agent <ua>` | Override Wikimedia User-Agent (default identifies this project) |
-| `--languages en,fr,...` | Explicitly narrow the default all-language sitelink set |
-| `--all-languages` | Explicit compatibility alias for the all-language default |
-| `--no-full-text` | Fetch only the lead section, not the full article |
-| `--max-articles-per-qid <n>` | Explicitly cap articles per QID (default: no cap) |
-| `--enrichment-batch-size <n>` | Maximum QIDs/titles per API batch (default `50`) |
-| `--enrichment-site-workers <n>` | Concurrent Wikipedia batch jobs (default `8`) |
-| `--limit <n>` | Process only the first N polygons per PBF |
-| `--skip-existing` | Skip PBFs already listed in the manifest |
-| `--force` | Re-process even when `--skip-existing` applies |
-| `--push` | Upload produced artifacts to the Hub |
-| `--upload-threads <n>` | Concurrent transfer workers in the atomic Hub commit (default `5`) |
-| `--commit-message <msg>` | Custom git commit message for the push |
-| `--dry-run` | Use a stub HF client (records calls without uploading) |
-| `--log-level <level>` | `DEBUG` / `INFO` / `WARNING` / `ERROR` |
-| `--dataset-version v1|v2` | Select the sync contract; V2 is opt-in and publishes to its separate Hub dataset |
-### Examples
-Process one PBF and write 3 parquet files + manifest locally:
-```bash
-uv run osm-polygon-wikidata-only process-pbf ~/pbfs/monaco-latest.osm.pbf
-```
-Push the result to the Hub with a stub client (no network):
-```bash
-uv run osm-polygon-wikidata-only process-pbf monaco-latest.osm.pbf --push --dry-run
-```
-Process every PBF under `<data-root>/raw/`, fetch only English and
-French Wikipedia, skip already-processed:
-```bash
-uv run osm-polygon-wikidata-only process-dir \
-    ~/pbfs/ \
-    --languages en,fr \
-    --skip-existing
-```
-Build the isolated V2 dataset. This keeps V1 artifacts intact, reuses their
-finalized documents and sidecars, and fetches only direct Wikipedia-tag pages
-that are not already present in V1:
-```bash
-uv run osm-polygon-wikidata-only sync-dir \
-    "$OSM_POLYGON_DATA_ROOT/raw" \
-    --dataset-version v2 \
-    --skip-existing \
-    --push
-```
-V2 local artifacts are written below `processed_v2/`; its request cache,
-resumable V1 reuse index, and restart checkpoints are below `cache/v2/`. The index is built one shard at
-a time while PBF extraction starts, and each Parquet row group is checkpointed
-across restarts. Its SQLite storage setup also runs in the background, so it
-does not delay extraction. A direct page that is not yet found may be fetched
-while the index continues. Once indexing finishes, a V1 match is preferred;
-only a page still absent from V1 keeps the fetched result. Both locations use
-the configured data root, so stopping and restarting the command is safe.
-Use `--force` only when deliberately rebuilding a V2 region.
-### Wikimedia Bot Password authentication
-Long production runs should authenticate their read-only API requests with a
-Wikimedia Bot Password. Authentication identifies the pipeline to Wikimedia
-and lets the account receive its applicable API request tier. It does not grant
-this project permission to edit Wikimedia.
-Create the credentials once:
-1. Sign in to the Wikimedia account that will operate the pipeline.
-2. Open [Meta-Wiki Special:BotPasswords](https://meta.wikimedia.org/wiki/Special:BotPasswords).
-3. Enter a descriptive bot name such as `osm-polygon-pipeline`.
-4. Select **Basic rights**, which are required for API reading. Leave editing,
-   page creation, uploading, and administration grants disabled; this pipeline
-   only reads public API data.
-5. Create the Bot Password and copy both generated values immediately. The
-   username includes a suffix, for example `AccountName@osm-polygon-pipeline`;
-   use the complete generated username, not the main account username.
-Export them into the current terminal session. Reading the password silently
-keeps it out of shell history:
-```bash
-export WIKIMEDIA_BOT_USERNAME='AccountName@osm-polygon-pipeline'
-read -rs WIKIMEDIA_BOT_PASSWORD
-export WIKIMEDIA_BOT_PASSWORD
-```
-Paste the generated password at the silent prompt and press Enter. Do not commit
-the password, add it to a checked-in `.env` file, paste it into an issue, or use
-the main Wikimedia account password. The process retains it only in memory.
-With both variables present, the unified pipeline uses the authenticated
-1,200-request-per-minute budget with one shared scheduler and at most eight requests in flight
-(anonymous runs default to three). To choose a different authenticated ceiling:
-```bash
-export WIKIMEDIA_REQUESTS_PER_MINUTE=600
-```
-Usually, omit this override and let the adaptive scheduler work. Wikimedia
-determines the actual tier from the account's standing: authentication alone
-does not guarantee a particular limit. The pipeline keeps at most eight requests
-in flight for authenticated runs, keeps cookies domain-scoped in one bounded persistent connection pool,
-and automatically reduces its rate when Wikimedia returns HTTP 429 with
-`Retry-After`.
-The startup log states either `anonymous` or `authenticated as <username>` and
-the configured ceiling. The password is never logged. Authentication itself is
-verified lazily on the first request to each Wikidata or language-Wikipedia host.
-If only one credential variable is set, configuration stops before processing;
-if Wikimedia rejects the pair, the first API request fails without silently
-falling back to anonymous access.
-Troubleshooting:
-- If the application names a missing variable, export both variables in the same
-  terminal that launches `uv run`.
-- If authentication is rejected, copy the complete generated bot username,
-  reset or recreate the Bot Password, and export the new values. Changing the
-  main account password can require resetting Bot Passwords.
-- If HTTP 429 responses continue, leave the automatic cooldown in control or
-  lower `WIKIMEDIA_REQUESTS_PER_MINUTE`; do not override the concurrency limit
-  without measuring the result.
-- To revoke access, return to
-  [Special:BotPasswords](https://meta.wikimedia.org/wiki/Special:BotPasswords)
-  and revoke the named Bot Password. Then remove the variables with
-  `unset WIKIMEDIA_BOT_USERNAME WIKIMEDIA_BOT_PASSWORD`.
-The implementation follows Wikimedia's official
-[Bot Password](https://www.mediawiki.org/wiki/Manual:Bot_passwords),
-[API login](https://www.mediawiki.org/wiki/API:Login), and
-[API rate-limit](https://www.mediawiki.org/wiki/Wikimedia_APIs/Rate_limits)
-guidance.
-### Resumable full-dataset command
-Run this single command to reconcile the existing augmentation backlog, repair
-remotely missing finalized artifacts without refetching them, process missing
-PBFs, immediately augment them, and publish complete regional bundles:
-```bash
-uv run osm-polygon-wikidata-only sync-dir "$OSM_POLYGON_DATA_ROOT/raw" \
-  --skip-existing \
-  --push
-```
-`process-dir` and `augment-dir` remain available as compatibility commands, but
-do not run them beside `sync-dir`. A data-root lock prevents duplicate unified
-runs. Each completed region is uploaded atomically with fresh manifests and the
-canonical dataset `README.md`. Internally the unified sync drains actions in
-this order: RECOVERY first for finalized shards eligible for an exhaustive
-local integrity audit (only affected QIDs are refetched and repaired transactionally),
-then AUGMENT backlog (each call performs Wikimedia sidecar work
-and may enqueue an atomic remote publication on success), then PUBLISH-only
-reconciliation repairs (Wikimedia-free -- each repair reuses the already-loaded
-local augmentation result and only enqueues a Hugging Face upload, with no
-extraction and no Wikidata / Wikipedia / Wikivoyage calls), then new core
-PROCESS work (the runner may prefetch the next PBF concurrently while
-enriching the current region), then COMPLETE / no-op states. Maps and the
-README are only reported "refreshed" after a successful core or metadata
-publication actually refreshed them and the background upload queue has
-drained.
-The recovery audit is resumable, content-addressed, and processed one region at
-a time. Healthy finalized regions store or reuse a receipt and continue; a
-damaged region is repaired and published immediately before the next region is
-audited. Stopping the command therefore preserves completed regional work,
-while changed inputs are checked again. A malformed or unreadable finalized
-shard fails closed instead of being silently skipped. Regions with incomplete
-sidecars are augmented first and audited immediately afterward in the same
-invocation. A repaired region is published atomically with its regional
-artifacts, manifests, and dataset card. Coverage maps are regenerated only
-when the repair changed one of their polygon, link, or document inputs.
-Legacy Wikidata fact rows whose subject is no longer present in the region's
-polygons are removed by that same transaction; all still-joinable facts are
-preserved.
-The command reports bounded local-scan and upstream-validation checkpoints with
-elapsed time. Up to three independent validation chunks overlap under the same
-shared request scheduler, while transient Wikimedia API states such as `maxlag`
-remain retryable and never become cached missing entities.
-Affected relationships are repaired in deterministic groups of 25 QIDs. Up to
-three independent groups run concurrently under the same shared global and
-per-host Wikimedia scheduler. Each completed group is stored immediately below
-`cache/wikidata_recovery/checkpoints/<stem>/<plan-hash>/`, so an interruption
-repeats only unfinished groups. The final regional Parquet files and manifests
-are still replaced atomically only after every group has completed.
-Known whole-file Geofabrik containment overlaps are retired safely during
-`sync-dir --push`: retained parents receive missing sidecar rows, contained
-child artifacts are removed remotely in one atomic commit, and local originals
-are preserved under `quarantine/containment-v1/`. A child PBF is ignored only
-after the durable retirement manifest records local preparation; a missing
-parent polygon blocks retirement rather than discarding data.
-To pause, stop the command with `Ctrl-C`. Run the identical command again to
-resume: completed PBFs remain skipped, while the interrupted PBF is retried
-because it has no completed manifest entry. The durable pending-publications
-manifest and the upload-queue state files persist across restarts, so a
-failed upload is always retryable on the next invocation. Stage timings are
-logged for every PBF. Tune large runs only when needed with
-`--enrichment-batch-size`, `--enrichment-site-workers`, and `--upload-threads`.
-The normal command fetches full text for every valid language-Wikipedia
-sitelink with no per-QID cap. If any expected article remains unresolved after
-retries, that PBF is not published; rerunning resumes from successful cache
-checkpoints. With `--push`, each locally complete PBF is queued for an atomic
-background upload while the next PBF starts processing. Shutdown waits for the
-queue, and unresolved uploads make the command exit nonzero and remain queued
-for the next invocation.
-Programmatic usage:
-```python
-from pathlib import Path
-from osm_polygon_wikidata_only.config.paths import DataRoot, resolve_data_root
-from osm_polygon_wikidata_only.config.settings import Settings
-from osm_polygon_wikidata_only.enrichment.wikipedia_client import HttpWikipediaClient
-from osm_polygon_wikidata_only.enrichment.wikidata_client import HttpWikidataClient
-from osm_polygon_wikidata_only.pipeline.processor import process_pbf
-data_root = resolve_data_root(repo_root=Path.cwd())
-data_root.ensure()
-settings = Settings(languages=("en", "fr"))
-wd = HttpWikidataClient(settings)
-wiki = HttpWikipediaClient(settings)
-result = process_pbf(
-    Path("monaco-latest.osm.pbf"),
-    data_root=data_root,
-    wikidata_client=wd,
-    wikipedia_client=wiki,
-    settings=settings,
-)
-print(result.polygon_count, "polygons")
-```
-## Reliability and performance behavior
-The pipeline is designed to preserve dataset completeness while keeping
-Wikimedia traffic polite:
-* Candidate order, selected sitelinks, and Parquet row ordering are
-  deterministic.
-* Identical Wikidata QIDs and Wikipedia titles are fetched once per run and
-  reused for every matching polygon.
-* HTTP clients use the on-disk cache by default. Failed requests are cached
-  briefly to avoid repeatedly hammering a failing endpoint.
-* Concrete HTTP clients batch compatible Wikidata and same-language Wikipedia
-  requests. The pipeline falls back to the established per-item request path
-  if a batch response is incomplete or invalid.
-* If a valid page returns an empty TextExtracts result, the client parses that
-  page's exact revision through the Action API and converts the rendered HTML
-  to plain text while preserving the original revision ID.
-* Per-host pacing, retries with jitter, and a shared `429` cooldown remain in
-  force when batch jobs run concurrently.
-* Long enrichment stages emit a concise two-minute heartbeat naming the active
-  Wikidata or Wikipedia phase, completed and total QIDs, completed and total
-  Wikipedia sites, and articles attempted. The snapshot confirms liveness; it is
-  not an ETA and does not change request pacing. Short enrichment stages finish
-  without a heartbeat.
-* `--push` publishes every produced Parquet artifact and the final manifest in
-  one atomic Hugging Face commit. Transfers use concurrent workers; increase
-  `--upload-threads` only when local bandwidth and Hub quotas allow it.
-For a repeatable production run, use `--skip-existing`; it consults the
-manifest and leaves previously completed PBFs untouched. Use `--force` only
-when you intentionally want to rebuild a completed PBF.
-## Development quality checks
-The project exposes the same uv-managed checks locally and in GitHub
-Actions through `just`:
-```bash
-just check
-```
-Install the pre-commit hooks once per checkout:
-```bash
-uv run pre-commit install
-```
-The test suite uses in-memory clients and stub PBF readers for unit coverage.
-It does not require a real PBF, external data root, or Wikimedia request.
-### Read-only remote audit
-The separate operator command uses Typer for its CLI, Rich for its report, and
-tqdm for interactive per-region progress:
-```bash
-uv run osm-polygon-wikidata-only-audit-remote \
-  --data-root "$OSM_POLYGON_DATA_ROOT"
-```
-It compares finalized local artifacts with the Hugging Face inventory and
-prints a reconciliation plan. It does not modify local files or the remote
-dataset. The production `osm-polygon-wikidata-only` command remains a stable,
-non-interactive argparse interface; `sync-dir` keeps its low-noise log
-heartbeats instead of terminal progress bars.
----
-## Output schema
-Each PBF produces polygon and link tables plus canonical Wikipedia
-documents and the derived text/fact tables. The schema
-definitions live in `osm_polygon_wikidata_only.domain.schema` and
-`osm_polygon_wikidata_only.augmentation.schema` so the dataset
-card, the parquet writers, and the tests share a single source of
-truth.
-### `polygons/<stem>.parquet`
-One row per polygon. Includes geometry metadata, OSM tags, primary
-OSM tag, area bucket, and Wikipedia coverage counters.
-### `wikipedia/documents/<stem>.parquet`
-One row per unique Wikipedia article
-(`(wikidata, language, page_id, revision_id)`). Includes lead text,
-plain-text full text, thumbnails, license, attribution, and a
-deterministic SHA-256 `content_hash`. It preserves every field from
-the former `articles/` table and adds stable `document_id` and
-`project` fields. The legacy remote `articles/` path is deleted in
-the same atomic Hub commit as the canonical upload; the local
-`processed/articles/` staging file is removed only after confirmed
-publication and reference validation.
-### `polygon_articles/<stem>.parquet`
-Unified many-to-many links joining polygons to Wikipedia and Wikivoyage
-documents. The `project` column identifies the source corpus and
-`document_id` references the corresponding project document table.
-### `manifests/processed_pbfs.json`
-Aggregate stats per source PBF: polygon/article counts, language
-coverage, area-bucket counts, top tag keys.
-### Text and fact tables
-Text sections, Wikivoyage documents, and Wikidata facts are published
-when augmentation has run for a region. Wikipedia sections reference
-the canonical document IDs; retiring the legacy article table never
-removes or rewrites section content.
-- `wikipedia/sections/<stem>.parquet` — section-level partitions of
-  Wikipedia document text.
-- `wikivoyage/documents/<stem>.parquet` — full Wikivoyage documents
-  associated with places through Wikidata.
-- `wikivoyage/sections/<stem>.parquet` — section-level partitions of
-  Wikivoyage document text.
-- `wikidata/facts/<stem>.parquet` — structured Wikidata claims for
-  polygon entities.
-Column lists for documents, sections, and facts live in
-`osm_polygon_wikidata_only.augmentation.schema` (`DOCUMENT_COLUMNS`,
-`SECTION_COLUMNS`, `FACT_COLUMNS`) and the documented column
-descriptions live in
-`osm_polygon_wikidata_only.augmentation.schema_descriptions`. The
-generated Hugging Face dataset card (`README.md`) embeds every
-augmentation column with its description; the dataset card renderer
-is the source of truth for what's published.
-## Generated dataset card
-The published dataset card on the Hugging Face Hub is regenerated
-automatically before every publication path. It reports factual
-core, Wikipedia, Wikivoyage, section, and Wikidata-fact statistics
-computed directly from the local finalized Parquet files under
-`<processed>/`. No hardcoded counts live in the source repository
-or the generated card; every figure is recomputed on each
-publication.
-The canonical renderer is `osm_polygon_wikidata_only.hf.publication.write_readme_snapshot`,
-backed by `osm_polygon_wikidata_only.hf.dataset_stats.render_stats_section`
-and `osm_polygon_wikidata_only.hf.dataset_card.render_dataset_card`.
-The renderer pulls both the core `DatasetStats` snapshot and the
-private augmentation snapshot from the local sidecars.
----
-## Geographic coverage
 ### Polygons with Wikipedia or Wikivoyage text
 ![Polygons with Wikipedia or Wikivoyage text](assets/geographic_text_presence.png)
-Each point is a dataset polygon with at least one non-empty Wikipedia
-document or a non-empty Wikivoyage document sharing its Wikidata entity.
-A polygon is shown once even when several documents qualify.
 ### All dataset polygons
 ![Coverage Map](assets/coverage_map.png)
-Each point represents one dataset polygon carrying an OSM `wikidata=*`
-tag, whether or not corresponding Wikipedia or Wikivoyage text exists.
 ### Wikipedia + Wikivoyage text density
 ![Geographic Wikipedia and Wikivoyage Text Density](assets/geographic_text_density.png)
-Each H3 cell contains the raw number of polygons with non-empty text
-from Wikipedia or Wikivoyage. A polygon is counted once even when both
-projects or several documents qualify. Colour uses a logarithmic
-purple-to-yellow scale; this is an absolute density count, not a proportion
-of all polygons.
----
-## Wikimedia etiquette
-Wikimedia APIs require a User-Agent identifying the project and a
-contact. The defaults are in `config.settings.DEFAULT_USER_AGENT`.
-Set `--user-agent` in production deployments.
-The HTTP clients honor:
-* configurable `request_timeout_s`, `request_max_retries`,
-  `request_base_delay_s`,
-* exponential backoff with jitter (`utils.retry.with_retries`),
-* a disk-backed `JsonFileCache` (`io.cache.JsonFileCache`) that lets
-  repeated runs avoid hammering the same endpoint,
-* localized language lists (`--languages`) so we never fetch
-  unwanted sitelinks.
-For authenticated production processing, follow the
-[Bot Password setup](#wikimedia-bot-password-authentication). Bot Password
-cookies are kept in memory and isolated per Wikimedia API host. Anonymous mode
-remains available when neither credential environment variable is set.
----
-## Development
-### Run the tests
-```bash
-just test
-```
-The 1,300+ tracked tests are deterministic and require no live network;
-HTTP clients come in three flavors (`Http…`, `InMemory…`,
-`Cached…`) and the tests use the in-memory flavors.
-### Lint and format
-```bash
-just lint
-just format
-```
-### Type-check
-```bash
-just typecheck
 ```
-Run `just --list` for every available recipe. All Python commands are executed
-through uv; GitHub Actions uses the same coverage, lint, format-check,
-type-check, and build recipes.
----
-## Repository / data separation policy
-The repository is **code-only**. Everything user-generated (datasets,
-HF caches, Arrow/Parquet files, downloaded PBFs) is git-ignored and
-must live on the configured external data root. This keeps the repo
-tiny, makes data updates cheap, and prevents accidental commits of
-multi-GB artifacts.
----
-## Licensing and attribution
-* **OpenStreetMap polygons**: (c) OpenStreetMap contributors, licensed
-  under [ODbL 1.0](https://opendatacommons.org/licenses/odbl/).
-* **Wikidata** entity data: (c) Wikimedia contributors under
-  [CC0 1.0](https://creativecommons.org/publicdomain/zero/1.0/).
-* **Wikipedia** article text: (c) respective Wikipedia authors,
-  licensed under
-  [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/).
-  Attribution and license are stored inline in the
-  `wikipedia/documents/<stem>.parquet` `license` and `attribution`
-  columns.
-Any derivative dataset must preserve OSM attribution as described on
-<https://www.openstreetmap.org/copyright>.

+---
+license: odbl
+language:
+  - en
+tags:
+  - openstreetmap
+  - wikidata
+  - wikipedia
+  - wikivoyage
+  - polygons
+  - geospatial
+  - multilingual
+configs:
+  - config_name: polygons
+    data_files:
+      - split: polygons
+        path: polygons/*.parquet
+  - config_name: polygon_articles
+    data_files:
+      - split: polygon_articles
+        path: polygon_articles/*.parquet
+  - config_name: wikipedia_documents
+    data_files:
+      - split: wikipedia_documents
+        path: wikipedia/documents/*.parquet
+  - config_name: wikipedia_sections
+    data_files:
+      - split: wikipedia_sections
+        path: wikipedia/sections/*.parquet
+  - config_name: wikivoyage_documents
+    data_files:
+      - split: wikivoyage_documents
+        path: wikivoyage/documents/*.parquet
+  - config_name: wikivoyage_sections
+    data_files:
+      - split: wikivoyage_sections
+        path: wikivoyage/sections/*.parquet
+  - config_name: wikidata_facts
+    data_files:
+      - split: wikidata_facts
+        path: wikidata/facts/*.parquet
+dataset_info:
+  polygon_count: 1184110
+  unique_wikidata_count: 1119223
+  article_count: 2273750
+---
+# NoeFlandre/osm-polygon-wikidata-only
+![NoeFlandre/osm-polygon-wikidata-only dataset overview](assets/dataset_hero.png)
+OSM polygons tagged with a `wikidata=*` reference, enriched with Wikipedia and Wikivoyage text across all available languages. The published tables are:
+- `polygons/<stem>.parquet` — one row per polygon
+- `wikipedia/documents/<stem>.parquet` — one row per unique Wikipedia article
+- `polygon_articles/<stem>.parquet` — unified polygon-to-document many-to-many links for Wikipedia and Wikivoyage; `project` identifies the source and `document_id` references its document table
+- `wikipedia/sections/<stem>.parquet` — section-level partitions of Wikipedia document text
+- `wikivoyage/documents/<stem>.parquet` — full Wikivoyage documents associated with places through Wikidata
+- `wikivoyage/sections/<stem>.parquet` — section-level partitions of Wikivoyage document text
+- `wikidata/facts/<stem>.parquet` — structured Wikidata claims for polygon entities
+Links are derived from the Wikidata identifiers shared by each OSM polygon and its Wikipedia or Wikivoyage documents.
+Generated on 2026-08-02.
+Maintained by **Noé Flandre**.
+Source code: [GitHub repository](https://github.com/NoeFlandre/osm-polygon-wikidata-only).
+This Hugging Face dataset is **V1**, produced by the [V1.0.0 GitHub code](https://github.com/NoeFlandre/osm-polygon-wikidata-only/tree/v1.0.0).
 ## Trackio snapshot
+[View the dataset presentation](https://noeflandre.github.io/osm-polygon-wikidata-only/presentations/dataset.html) for a visual overview of the published tables and geographic coverage.
+The finished dataset is recorded in the single public Trackio run [`final-dataset-snapshot`](https://huggingface.co/spaces/NoeFlandre/osm-polygon-wikidata-only-trackio). It contains exactly three plots: a text-coverage funnel, the top ten Wikipedia languages plus `Other languages`, and dataset composition on a logarithmic scale. The run is a static snapshot with no pipeline timeline.
 | Snapshot metric | Value |
 | --- | ---: |
 | Polygon/link-table storage | 9.9 GB |
 | Total Parquet storage | 19.2 GB |
+The funnel's language thresholds use the canonical Wikipedia polygon fields. Wikivoyage is included in the combined document, word, and dataset-composition totals.
+## Coverage
 ### Polygons with Wikipedia or Wikivoyage text
 ![Polygons with Wikipedia or Wikivoyage text](assets/geographic_text_presence.png)
+Each point is a dataset polygon with at least one non-empty Wikipedia document or a non-empty Wikivoyage document sharing its Wikidata entity. A polygon is shown once even when several documents qualify.
 ### All dataset polygons
 ![Coverage Map](assets/coverage_map.png)
+Each point represents one dataset polygon carrying an OSM `wikidata=*` tag, whether or not corresponding Wikipedia or Wikivoyage text is available.
+## Geographic coverage
 ### Wikipedia + Wikivoyage text density
 ![Geographic Wikipedia and Wikivoyage Text Density](assets/geographic_text_density.png)
+Each H3 cell contains the raw number of polygons with non-empty Wikipedia or Wikivoyage text. A polygon is counted once even when both projects or several documents qualify. Colour uses a logarithmic purple-to-yellow scale; this is an absolute density count, not a proportion of all polygons.
+## Dataset snapshot
+| Metric | Value |
+| --- | ---: |
+| Polygons | 1,184,110 |
+| Unique Wikidata entities | 1,119,223 |
+| Wikipedia polygon-document links | 2,468,604 |
+| Wikipedia + Wikivoyage languages | 351 |
+| Geographic regions | 375 |
+| Polygon and link tables size | 9.9 GB |
+| Wikipedia documents | 2,273,750 |
+| Wikipedia sections | 11,997,165 |
+| Wikivoyage documents | 14,420 |
+| Wikivoyage sections | 302,200 |
+| Wikidata facts | 3,901,092 |
+| Wikipedia + Wikivoyage document words | 801,528,334 |
+| Total Parquet size | 19.2 GB |
+Wikipedia + Wikivoyage document words sums the full Wikipedia and Wikivoyage documents and excludes section rows because sections duplicate document text.
+## Language distribution
+Top 20 languages across Wikipedia and Wikivoyage documents:
+Document counts include canonical rows from both projects. Polygon counts deduplicate places per language and require non-empty text: Wikipedia uses the `polygon_articles` link table, while Wikivoyage joins through the shared Wikidata QID.
+| Language | Documents | % of total | Polygons with non-empty text |
+| --- | ---: | ---: | ---: |
+| en | 223,301 | 10.1% | 238,950 |
+| de | 146,312 | 6.6% | 158,550 |
+| fr | 109,720 | 5.0% | 120,451 |
+| ceb | 105,479 | 4.8% | 113,410 |
+| sv | 74,518 | 3.4% | 79,076 |
+| ru | 73,228 | 3.3% | 80,023 |
+| es | 70,808 | 3.2% | 78,504 |
+| it | 65,927 | 3.0% | 74,554 |
+| pl | 65,685 | 3.0% | 70,164 |
+| zh | 59,285 | 2.7% | 64,758 |
+| ja | 54,072 | 2.5% | 58,226 |
+| arz | 53,882 | 2.4% | 57,236 |
+| nl | 52,953 | 2.4% | 58,130 |
+| uk | 50,059 | 2.3% | 56,242 |
+| fa | 38,840 | 1.8% | 42,279 |
+| pt | 36,022 | 1.6% | 40,519 |
+| ce | 32,849 | 1.5% | 34,597 |
+| ca | 32,520 | 1.5% | 36,286 |
+| cs | 28,497 | 1.3% | 32,286 |
+| no | 28,424 | 1.3% | 31,793 |
+**Concentration:**
+- Top 1 language: 10.1% of all Wikipedia + Wikivoyage documents
+- Top 5 languages: 29.9% of all Wikipedia + Wikivoyage documents
+- Top 10 languages: 45.1% of all Wikipedia + Wikivoyage documents
+- Top 20 languages: 63.6% of all Wikipedia + Wikivoyage documents
+**Long-tail:**
+- 351 language(s) total
+- 5 language(s) appear in fewer than 5 Wikipedia + Wikivoyage documents
+- 5 language(s) appear in fewer than 10 Wikipedia + Wikivoyage documents
+- 4 language(s) appear in fewer than 5 polygons with non-empty text
+## Storage accounting
+Core tables size and total storage size are additive:
+| Metric | Bytes | Human-readable |
+| --- | ---: | --- |
+| Polygon and link tables size | 2,243,135,062 | 2.1 GB |
+| Wikipedia, Wikivoyage, and Wikidata tables size | 18,338,737,425 | 17.1 GB |
+| Total Parquet size | 20,581,872,487 | 19.2 GB |
+## Wikipedia text corpus
+### Documents
+| Metric | Value |
+| --- | ---: |
+| Document rows | 2,273,750 |
+| Unique documents | 2,192,426 |
+| Unique Wikidata entities | 625,432 |
+| Languages | 351 |
+| Non-empty text | 2,273,693 |
+| Empty / null text | 57 |
+| Characters | 5,324,562,350 |
+| Words | 783,679,126 |
+| Estimated tokens | 1,330,287,296 |
+| Regions / files represented | 375 |
+Estimated tokens use a transparent approximation: `characters // 4` per non-empty text row, with a minimum of one token per non-empty row.
+### Sections
+| Metric | Value |
+| --- | ---: |
+| Section rows | 11,997,165 |
+| Unique sections | 11,549,631 |
+| Documents represented | 2,189,484 |
+| Avg sections per represented document | 5.48 |
+| Non-empty section rate | 100.0% |
+| Unique Wikidata entities | 625,417 |
+| Languages | 351 |
+| Non-empty text | 11,997,165 |
+| Empty / null text | 0 |
+| Characters | 7,066,582,723 |
+| Words | 1,108,894,757 |
+| Estimated tokens | 1,762,125,212 |
+| Regions / files represented | 375 |
+Estimated tokens use a transparent approximation: `characters // 4` per non-empty text row, with a minimum of one token per non-empty row.
+### Languages
+| Language | Documents | % of total |
+| --- | ---: | ---: |
+| en | 224,462 | 9.9% |
+| de | 149,091 | 6.6% |
+| fr | 112,223 | 4.9% |
+| ceb | 110,289 | 4.9% |
+| sv | 76,085 | 3.3% |
+| ru | 75,557 | 3.3% |
+| es | 72,538 | 3.2% |
+| it | 69,103 | 3.0% |
+| pl | 66,276 | 2.9% |
+| zh | 60,408 | 2.7% |
+## Wikivoyage text corpus
+### Documents
+| Metric | Value |
+| --- | ---: |
+| Document rows | 14,420 |
+| Unique documents | 14,137 |
+| Unique Wikidata entities | 5,361 |
+| Languages | 27 |
+| Non-empty text | 14,417 |
+| Empty / null text | 3 |
+| Characters | 112,794,734 |
+| Words | 17,849,208 |
+| Estimated tokens | 28,193,232 |
+| Regions / files represented | 375 |
+Estimated tokens use a transparent approximation: `characters // 4` per non-empty text row, with a minimum of one token per non-empty row.
+### Sections
+| Metric | Value |
+| --- | ---: |
+| Section rows | 302,200 |
+| Unique sections | 296,934 |
+| Documents represented | 14,135 |
+| Avg sections per represented document | 21.38 |
+| Non-empty section rate | 100.0% |
+| Unique Wikidata entities | 5,360 |
+| Languages | 27 |
+| Non-empty text | 302,200 |
+| Empty / null text | 0 |
+| Characters | 132,242,637 |
+| Words | 21,741,293 |
+| Estimated tokens | 32,946,346 |
+| Regions / files represented | 375 |
+Estimated tokens use a transparent approximation: `characters // 4` per non-empty text row, with a minimum of one token per non-empty row.
+### Languages
+| Language | Documents | % of total |
+| --- | ---: | ---: |
+| en | 3,434 | 23.8% |
+| de | 1,898 | 13.2% |
+| pl | 1,315 | 9.1% |
+| it | 1,043 | 7.2% |
+| fa | 970 | 6.7% |
+| fr | 900 | 6.2% |
+| zh | 537 | 3.7% |
+| ru | 535 | 3.7% |
+| es | 513 | 3.6% |
+| nl | 509 | 3.5% |
+## Wikidata facts
+| Metric | Value |
+| --- | ---: |
+| Fact rows | 3,901,092 |
+| Unique facts | 3,836,531 |
+| Unique subjects | 955,700 |
+| Distinct properties | 8 |
+| Non-empty English property label | 3,901,092 |
+| Non-empty English value label | 3,535,905 |
+| With qualifiers | 232,610 |
+| With references | 1,756,783 |
+| Unreadable qualifier JSON | 3,668,482 |
+| Unreadable references JSON | 2,144,309 |
+| Regions / files represented | 375 |
+**Value-type distribution:**
+| Value type | Facts |
+| --- | ---: |
+| wikibase-item | 3,535,905 |
+| time | 267,269 |
+| quantity | 97,918 |
+**Top properties:**
+| Property ID | English label | Facts |
+| --- | --- | ---: |
+| P31 | instance of | 1,166,284 |
+| P131 | located in the administrative territorial entity | 985,587 |
+| P17 | country | 984,654 |
+| P1435 | heritage designation | 336,569 |
+| P571 | inception | 267,269 |
+| P2044 | elevation above sea level | 97,918 |
+| P361 | part of | 60,996 |
+| P279 | subclass of | 1,815 |
+English labels are requested where available; multilingual labels are preserved verbatim in the Parquet column `property_labels`.
+## Geographic distribution by continent
+This table is recomputed from the finalized Parquet tables before each dataset-card publication. Each polygon's WGS84 centroid is spatially matched to the bundled Natural Earth 1:110m Admin-0 country boundaries, then assigned the country's continent. Because this is a coarse global reference, offshore points and centroids outside its country polygons remain `Unassigned` rather than being guessed.
+| Continent | Polygons | Wikipedia documents | Wikivoyage documents | Polygons with Wikipedia text | Polygons with Wikipedia or Wikivoyage text | Text coverage |
+| --- | ---: | ---: | ---: | ---: | ---: | ---: |
+| Africa | 11,289 | 55,254 | 550 | 8,309 | 8,309 | 73.6% |
+| Antarctica | 254 | 1,684 | 1 | 251 | 251 | 98.8% |
+| Asia | 117,610 | 387,567 | 1,811 | 94,924 | 94,945 | 80.7% |
+| Europe | 792,569 | 1,073,709 | 5,163 | 373,782 | 373,782 | 47.2% |
+| North America | 162,012 | 392,739 | 3,429 | 120,166 | 120,167 | 74.2% |
+| Oceania | 19,807 | 30,678 | 396 | 11,467 | 11,468 | 57.9% |
+| South America | 18,906 | 63,181 | 786 | 13,503 | 13,503 | 71.4% |
+| Unassigned | 61,663 | 193,785 | 2,007 | 43,824 | 43,826 | 71.1% |
+**Metric definitions:**
+- `Polygons`: dataset polygons whose centroid is assigned to the continent. Every polygon appears in exactly one continent row, including `Unassigned`.
+- `Wikipedia documents`: distinct non-empty Wikipedia documents connected to those polygons through `polygon_articles`. A document is counted once within a continent, but may appear in more than one continent when linked polygons span more than one continent.
+- `Wikivoyage documents`: distinct non-empty Wikivoyage documents whose Wikidata entity is shared by a polygon in the continent. The same cross-continent counting rule applies.
+- `Polygons with Wikipedia text`: polygons linked to at least one non-empty Wikipedia document.
+- `Polygons with Wikipedia or Wikivoyage text`: polygons satisfying the Wikipedia condition or sharing a Wikidata entity with at least one non-empty Wikivoyage document. Each polygon is counted once.
+- `Text coverage`: `combined text-covered polygons / all dataset polygons` in the continent row.
+## Schema
+### `polygons`
+| Column | Description |
+| --- | --- |
+| `polygon_id` | Deterministic ID: `<source_pbf_stem>:<osm_type>:<osm_id>`. |
+| `region` | Geofabrik region slug parsed from the source PBF filename (e.g. `monaco`). |
+| `source_pbf` | Source PBF filename (e.g. `monaco-latest.osm.pbf`). |
+| `osm_type` | OSM element type: `way` or `relation`. |
+| `osm_id` | OpenStreetMap numeric identifier of the element. |
+| `wikidata` | Wikidata Q-id from the OSM `wikidata=*` tag. |
+| `name` | Convenience: `tags.name` if present, empty string otherwise. |
+| `tags` | Deterministic JSON object of all OSM tags except `wikidata`. |
+| `tag_keys` | Deterministic JSON list of sorted OSM tag keys. |
+| `tag_count` | Number of OSM tags (excluding `wikidata`). |
+| `osm_primary_tag` | Best single tag for coarse analysis, e.g. `landuse=forest`. |
+| `centroid` | Polygon centroid as a GeoJSON Point string (`[lon, lat]`). |
+| `lat` | Centroid latitude in decimal degrees (WGS84). |
+| `lon` | Centroid longitude in decimal degrees (WGS84). |
+| `bbox` | Bounding box as JSON list `[min_lon, min_lat, max_lon, max_lat]`. |
+| `geometry` | Full polygon geometry as deterministic GeoJSON string in WGS84 lon/lat coordinates. |
+| `area_m2` | Polygon area in square meters (WGS84 equirectangular approximation). |
+| `area_km2` | Polygon area in square kilometers. |
+| `area_bucket` | Human-readable size bucket (e.g. `1-10km2`). |
+| `has_name` | True if the polygon has a `name` tag. |
+| `has_wikidata` | True if the polygon has a `wikidata` tag (always true by filter). |
+| `has_wikipedia` | True if at least one linked Wikipedia article was fetched. |
+| `wikipedia_language_count` | Number of Wikipedia languages for which articles were linked. |
+| `wikipedia_languages` | Deterministic JSON list of language codes (e.g. `["en","fr"]`). |
+| `wikipedia_article_count` | Number of unique article revisions linked to this polygon. |
+| `has_english_wikipedia` | True if `en` is among the available languages. |
+| `has_french_wikipedia` | True if `fr` is among the available languages. |
+| `text_available` | True if at least one linked article has non-empty full text. |
+| `best_language` | Deterministic preferred language code (e.g. `en`). |
+| `extraction_version` | Package version that produced the row. |
+| `extracted_at` | ISO-8601 UTC timestamp at the moment the row was extracted. |
+### `wikipedia/documents`
+| Column | Description |
+| --- | --- |
+| `document_id` | Deterministic document identifier (`<wikidata>:<project>:<language>:<page_id>:<revision_id>`). |
+| `article_id` | Deterministic ID: `<wikidata>:<language>:<page_id>:<revision_id>`. |
+| `wikidata` | Wikidata Q-id this article is linked to. |
+| `project` | Wiki project name: always `wikipedia` for this table. |
+| `language` | Wikipedia language code, e.g. `en`. |
+| `site` | Wikidata sitelink site, e.g. `enwiki`. |
+| `title` | Article title as returned by the Wikipedia API. |
+| `url` | Canonical Wikipedia article URL. |
+| `page_id` | MediaWiki page ID. |
+| `revision_id` | Exact revision ID used for text extraction. |
+| `revision_timestamp` | ISO-8601 timestamp of the revision. |
+| `retrieved_at` | ISO-8601 UTC timestamp when this pipeline fetched the article. |
+| `wikidata_label` | Best Wikidata label for the article's language, fallback English. |
+| `wikidata_description` | Best Wikidata description for the article's language. |
+| `wikidata_aliases` | Deterministic JSON list of Wikidata aliases. |
+| `lead_text` | Lead section of the article, plain text. |
+| `extract` | Short summary/extract if the API provided one. |
+| `full_text` | Full cleaned article text, plain text. |
+| `full_text_format` | Encoding of `full_text`; always `plain_text`. |
+| `article_length_chars` | Length of `full_text` in characters. |
+| `article_length_words` | Approximate whitespace-token count of `full_text`. |
+| `article_length_tokens_estimate` | Rough token count: `chars / 4`. |
+| `thumbnail_url` | Thumbnail URL only (no image bytes stored). |
+| `thumbnail_width` | Thumbnail width in pixels, if known. |
+| `thumbnail_height` | Thumbnail height in pixels, if known. |
+| `categories` | Deterministic JSON list of category titles, or `[]`. |
+| `license` | License string, e.g. `CC BY-SA`. |
+| `attribution` | Attribution string for the article. |
+| `source_api` | Which API was queried: `mediawiki_action_api` / `wikipedia_rest_api`. |
+| `fetch_status` | One of: `ok`, `article_not_found`, `http_error`, `rate_limited`, `parse_error`, `empty_text`. |
+| `fetch_error` | Short diagnostic on failure, empty string on success. |
+| `content_hash` | Stable SHA-256 of `full_text` for change tracking. |
+### `polygon_articles`
+| Column | Description |
+| --- | --- |
+| `polygon_id` | FK to polygons polygon_id. |
+| `document_id` | FK to the project documents table. Format: <wikidata>:<project>:<language>:<page_id>:<revision_id>. |
+| `project` | Source project: wikipedia or wikivoyage. |
+| `wikidata` | Wikidata QID (denormalized for fast filtering). |
+| `language` | Document language code. |
+| `source_pbf` | Source PBF filename (denormalized). |
+| `region` | Geofabrik region slug (denormalized). |
+| `osm_type` | OSM element type: way or relation. |
+| `osm_id` | OSM numeric identifier. |
+| `page_id` | MediaWiki page ID. |
+| `revision_id` | MediaWiki revision ID. |
+### `wikivoyage/documents`
+| Column | Description |
+| --- | --- |
+| `document_id` | Deterministic document identifier (`<wikidata>:<project>:<language>:<page_id>:<revision_id>`). |
+| `article_id` | Stable article identifier referenced by `polygon_articles/<stem>.parquet`. |
+| `wikidata` | Wikidata QID this document is linked to. |
+| `project` | Wiki project name: `wikipedia` or `wikivoyage`. |
+| `language` | Wikipedia or Wikivoyage language code (e.g. `en`). |
+| `site` | Wikidata sitelink host, e.g. `enwiki` or `enwikivoyage`. |
+| `title` | Page title as returned by the MediaWiki API. |
+| `url` | Canonical URL of the page. |
+| `page_id` | MediaWiki page ID (integer). |
+| `revision_id` | MediaWiki revision ID used to fetch the page text (integer). |
+| `revision_timestamp` | ISO-8601 timestamp of the revision. |
+| `retrieved_at` | ISO-8601 UTC timestamp when the pipeline fetched the page. |
+| `full_text` | Cleaned plain-text document body (Wikipedia articles or Wikivoyage pages). |
+| `full_text_format` | Encoding of `full_text`; always `plain_text`. |
+| `article_length_chars` | Length of `full_text` in characters. |
+| `article_length_words` | Approximate whitespace-token count of `full_text`. |
+| `article_length_tokens_estimate` | Rough token estimate: `chars / 4`. |
+| `license` | License string (`CC BY-SA` for Wikipedia/Wikivoyage text). |
+| `attribution` | Attribution string for the page. |
+| `source_api` | Which API was queried: `mediawiki_action_api` or `wikivoyage_rest_api`. |
+| `fetch_status` | One of: `ok`, `page_not_found`, `http_error`, `rate_limited`, `parse_error`, `empty_text`. |
+| `fetch_error` | Short diagnostic on failure, empty string on success. |
+| `content_hash` | Stable SHA-256 of `full_text` for change tracking. |
+### `wikipedia/sections` and `wikivoyage/sections`
+| Column | Description |
+| --- | --- |
+| `section_id` | Deterministic SHA-256 over `(document_id, section_index, anchor)`. |
+| `document_id` | FK back to `documents` (Wikipedia or Wikivoyage). |
+| `article_id` | FK back to the corresponding `articles` row. |
+| `wikidata` | Wikidata QID (denormalized for fast filtering). |
+| `project` | Wiki project name: `wikipedia` or `wikivoyage`. |
+| `language` | Wikipedia or Wikivoyage language code. |
+| `site` | Wikidata sitelink host. |
+| `page_id` | MediaWiki page ID (integer). |
+| `revision_id` | MediaWiki revision ID (integer). |
+| `section_index` | Sequential position of the section inside the document (integer). |
+| `heading` | Section heading, or empty string when the section is the lead. |
+| `anchor` | Section anchor after MediaWiki parsing. |
+| `level` | Heading level (1..6), or 0 for the lead section (integer). |
+| `parent_section_id` | Section ID of the enclosing section, or empty string. |
+| `section_path` | JSON array of ancestor section IDs, in order. |
+| `text` | Plain-text section body. |
+| `text_length_chars` | Length of `text` in characters (integer). |
+| `text_length_words` | Approximate whitespace-token count of `text` (integer). |
+| `text_length_tokens_estimate` | Rough token estimate: `chars / 4` (integer). |
+| `content_hash` | Stable SHA-256 of `section.text`. |
+| `license` | License string for this section. |
+| `attribution` | Attribution string for this section. |
+### `wikidata/facts`
+| Column | Description |
+| --- | --- |
+| `fact_id` | Deterministic SHA-256 over `(subject, property, value, ordinal)`. |
+| `wikidata` | Wikidata QID the fact belongs to (the subject). |
+| `property_id` | Property P-id (e.g. `P17`). |
+| `property_label_en` | English label for the property, when available. |
+| `property_labels` | Deterministic JSON object of property labels per language. |
+| `value_type` | Wikidata value datatype: `wikibase-entityid`, `string`, `quantity`, `time`, ... |
+| `value_entity_id` | Entity-valued object QID, when the value is a Wikidata entity. |
+| `value_label_en` | English label for the value entity, when available. |
+| `value_labels` | Deterministic JSON object of value labels per language. |
+| `value_text` | Rendered text representation of the value. |
+| `numeric_value` | Numeric amount for `quantity`-typed values, otherwise null. |
+| `unit_entity_id` | Wikidata QID of the unit (e.g. `Q11573` for metre). |
+| `rank` | Wikidata rank: `preferred`, `normal`, or `deprecated`. |
+| `qualifiers` | Deterministic JSON object of qualifier snaks, or `{}` when absent. |
+| `references` | Deterministic JSON array of reference groups, or `[]` when absent. |
+| `retrieved_at` | ISO-8601 UTC timestamp when the pipeline fetched the entity. |
+| `source_api` | Which API was queried (always `wikidata_action_api`). |
+## Data sources & licenses
+- **OpenStreetMap** polygons: (c) OpenStreetMap contributors, licensed under [ODbL 1.0](https://opendatacommons.org/licenses/odbl/).
+- **Wikidata** entity data: [CC0 1.0](https://creativecommons.org/publicdomain/zero/1.0/).
+- **Wikipedia** article text: licensed under [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/) by the respective Wikipedia editors; attributed inline per article.
+- **Wikivoyage** text: licensed under [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/) by the respective Wikivoyage editors; attributed inline per document.
+- **Natural Earth** Admin-0 geography: public-domain 1:110m reference data used only to assign centroid-based continent statistics and draw context maps.
+## How to load
+```python
+from datasets import load_dataset
+ds = load_dataset("parquet", data_files={
+    "polygons": "hf://datasets/NoeFlandre/osm-polygon-wikidata-only/polygons/*.parquet",
+})
 ```