Replace project README with dedicated dataset card

README.md

@@ -27,140 +27,72 @@ configs:
 
 # BNE Linked Data
 
-**Published dataset:** [hsilvosa/bne-linked-data on Hugging Face](https://huggingface.co/datasets/hsilvosa/bne-linked-data)
+## Dataset description
 
-This directory contains a reproducible pipeline for downloading, normalizing, profiling, and publishing the Linked Open Data dumps from the Biblioteca Nacional de España (BNE). The source represents bibliographic resources, authority records, and subject headings as an RDF knowledge graph.
+BNE Linked Data is a normalized Parquet release of the Linked Open Data dumps published by the Biblioteca Nacional de España. It represents bibliographic resources, authority records, subject headings, and the relationships among them as RDF statements.
 
-The dataset is observational and relational. It has no target label, predefined classes, or official train, validation, and test split. Each collection is published as a `train` split so downstream users can build graph, retrieval, entity-linking, or recommendation tasks without an arbitrary split imposed upstream.
+The released snapshot contains 260,288,924 triples from official BNE files last modified on 6 September 2023. The files were preserved by Internet Archive on 20 September 2023. Each RDF statement is represented by one row; no relationships are inferred and no enrichment is added.
 
 ## Source and licence
 
-- BNE Linked Data portal: <https://datos.bne.es/>
-- Official documentation and downloads: <https://www.bne.es/es/catalogos/datos-enlazados-bne/datos-enlazados>
-- Publisher: Biblioteca Nacional de España
-- Data licence: [CC0 1.0 Universal](https://creativecommons.org/publicdomain/zero/1.0/)
-- Released source snapshot: 6 September 2023, preserved by Internet Archive on 20 September 2023
+- Publisher: [Biblioteca Nacional de España](https://datos.bne.es/)
+- Documentation: [BNE Linked Data](https://www.bne.es/es/catalogos/datos-enlazados-bne/datos-enlazados)
+- Snapshot date: 6 September 2023
+- Licence: [CC0 1.0 Universal](https://creativecommons.org/publicdomain/zero/1.0/)
 
-The BNE documentation page describes an older 21 January 2021 extraction with more than four million bibliographic records, more than four million authority records, and approximately thirty million RDF triples. The reproducible release in this directory uses the later September 2023 official files preserved by Internet Archive. Direct source metadata identifies 6 September 2023 as the last-modified date for all three files.
+The live BNE download service returned HTTP 403 to automated requests when this release was built. The data files therefore come from preserved copies of the original BNE BZip2 responses in Internet Archive. Source URLs, dates, and transformations are recorded in `artifacts/provenance.json`.
 
-As checked on 12 August 2026, the live BNE download URLs and SPARQL endpoint return HTTP 403 to automated requests from the current execution network because of a Cloudflare rule. `configs/default.json` therefore points to byte-for-byte captures of the official BZip2 responses in Internet Archive. Downloads remain resumable, and `configs/archive-2023.json` records the same immutable sources explicitly.
+## Configurations
 
-## Dataset contents
-
-The release is designed around three Hugging Face configurations:
-
-| Configuration | Source content | Compressed source | RDF triples | Parquet |
+| Configuration | Content | Rows | Parquet files | Parquet size |
 |---|---|---:|---:|---:|
-| `bibliographic` | Bibliographic resources covering books, manuscripts, serials, maps, photographs, music, sound, and audiovisual material | 1,212,541,852 bytes | 208,059,171 | 1,772,292,848 bytes |
-| `authorities` | People, organisations, conferences, works, expressions, geographic names, and other authority entities | 359,784,392 bytes | 46,107,241 | 493,978,094 bytes |
-| `subjects` | Subject headings represented in SKOS | 33,805,185 bytes | 6,122,512 | 51,987,039 bytes |
+| `bibliographic` | Bibliographic resources, manifestations, works, and their relationships | 208,059,171 | 833 | 1,772,292,848 bytes |
+| `authorities` | People, organisations, conferences, works, expressions, and other authority entities | 46,107,241 | 185 | 493,978,094 bytes |
+| `subjects` | Subject headings and SKOS relationships | 6,122,512 | 25 | 51,987,039 bytes |
+
+Each configuration has a single `train` split containing the complete graph collection. The source does not define a supervised prediction target or official train, validation, and test partitions.
 
-Each RDF statement becomes one Parquet row. No triples are inferred, discarded, or enriched. The normalized columns are:
+## Fields
 
-| Column | Meaning |
+| Field | Description |
 |---|---|
-| `record_id` | Stable identifier formed from collection and decompressed source-line number |
+| `record_id` | Stable identifier formed from the collection and source-line number |
 | `subject` | Subject IRI or blank-node identifier |
 | `predicate` | Predicate IRI |
 | `object` | Object IRI, blank-node identifier, or decoded literal value |
 | `object_kind` | `iri`, `literal`, or `blank_node` |
 | `language` | Language tag attached to a literal, when present |
 | `datatype` | Datatype IRI attached to a literal, when present |
-| `collection` | Source collection name |
-| `source_line` | One-based line number in the decompressed dump |
-| `snapshot_date` | Date assigned by the source to the dump |
-
-The statement representation preserves the complete graph while remaining easy to query with DuckDB, Polars, Spark, or Hugging Face Datasets. Subjects and objects can be joined directly to reconstruct authorship, subject, work, expression, manifestation, and external authority relationships.
-
-## Capacity and duration
-
-The verified download is 1,606,131,429 bytes (1.50 GiB) and took 10 minutes 46 seconds on the measured connection. Normalization took 1 hour 39 minutes, the full DuckDB profile took 49 seconds, and the resulting 1,043 Parquet shards occupy 2,318,257,981 bytes (2.16 GiB). Allow at least 10 GiB free when both processed data and a staging copy are retained. The pipeline streams BZip2 input and does not create a decompressed N-Triples copy.
+| `collection` | Source graph collection |
+| `source_line` | One-based physical line number in the decompressed source |
+| `snapshot_date` | Source snapshot date |
 
 ## Exploratory profile
 
-The September 2023 snapshot contains 260,288,924 triples. Authorities have 5,624,909 distinct subjects and 96 predicates; bibliographic data has 17,850,587 distinct subjects and 105 predicates; subjects have 706,951 distinct subjects and 21 predicates. Objects are split between 85,517,669 IRIs and 174,771,255 literals. No blank-node objects occur in the snapshot.
-
-Subject labels contain 1,604,451 language-tagged literals: 1,527,636 Spanish, 32,508 English, 20,530 French, 15,982 Catalan, 5,186 Basque, and 2,609 Galician. Frequent relations include `rdf:type`, BNE identifiers and labels, bibliographic manifestation/work links, SKOS semantic relations, and 1,353,732 `owl:sameAs` authority links. Complete counts and the thirty most frequent predicates per collection are in `artifacts/profile.json`.
-
-Run the estimate before every download:
-
-```powershell
-bne-pipeline estimate --config configs/default.json
-```
-
-## Installation
-
-Python 3.11 or newer is recommended.
-
-```powershell
-python -m venv .venv
-.\.venv\Scripts\python.exe -m pip install -e ".[dev]"
-```
-
-## Pipeline
-
-Run each stage independently:
-
-```powershell
-bne-pipeline download --config configs/default.json
-bne-pipeline normalize --config configs/default.json
-bne-pipeline analyze --config configs/default.json
-bne-pipeline stage --config configs/default.json
-```
-
-Or execute the complete workflow:
-
-```powershell
-bne-pipeline run --config configs/default.json
-```
-
-Interrupted downloads are retained as `.part` files and resumed with an HTTP range request when the server supports it. A file is promoted to the raw source path only after its byte size matches the length announced by the server. The sizes recorded in configuration are rounded planning estimates published by BNE.
-
-If Cloudflare permits the files in a normal browser but blocks the pipeline, download them from the official documentation and place them at these exact paths:
-
-```text
-data/raw/autoridades.nt.bz2
-data/raw/materias.nt.bz2
-data/raw/bibliograficos.nt.bz2
-```
-
-Then start at `normalize`. Existing sources are checked for a non-empty BZip2 header, so an HTML block page cannot be mistaken for an RDF dump. Full BZip2 integrity and UTF-8/N-Triples validity are checked while the files are streamed during normalization.
-
-The September 2023 subjects dump contains a small number of external BnF IRIs split immediately before the closing `>` and continued on the next physical line. The normalizer rejoins only this identifiable two-line pattern, preserves the first physical line in `source_line`, and reports the number of repaired statements. It does not apply general-purpose cleanup or alter IRI text.
-
-## Small real-data validation
+The authority collection contains 5,624,909 distinct subjects and 96 predicates. The bibliographic collection contains 17,850,587 distinct subjects and 105 predicates. The subject collection contains 706,951 distinct subjects and 21 predicates.
 
-`configs/sample.json` uses the smallest archived official 2023 dump, subject headings, and stops normalization after 10,000 valid triples. Its purpose is to validate the real RDF structure without producing a complete release. The 33.8 MB BZip2 file still has to be downloaded in full because BZip2 cannot be decoded safely from an arbitrary byte range.
+Across all configurations there are 85,517,669 IRI objects and 174,771,255 literal objects. No blank-node objects occur. Subject data contains 1,604,451 language-tagged literals: 1,527,636 Spanish, 32,508 English, 20,530 French, 15,982 Catalan, 5,186 Basque, and 2,609 Galician.
 
-```powershell
-bne-pipeline estimate --config configs/sample.json
-bne-pipeline run --config configs/sample.json
-```
+Frequent relations include `rdf:type`, BNE identifiers and labels, bibliographic work and manifestation links, SKOS semantic relations, and 1,353,732 `owl:sameAs` authority links. Full counts and the thirty most frequent predicates per configuration are available in `artifacts/profile.json`.
 
-The complete September 2023 snapshot uses `configs/default.json`; sample outputs are not mixed into a final release.
+## Processing and quality
 
-When the source returns HTTP 403, `configs/fixture.json` provides an offline integration check using five real `owl:sameAs` statements for the BNE authority record of Miguel de Cervantes (`XX1718747`). These links to DNB, VIAF, DBpedia, IdRef, and Libris are reproduced from the [MARiMbA project documentation](https://oeg.fi.upm.es/index.php/es/technologies/228-marimba/index.html) and the [BNE/OEG project presentation](https://datos.gob.es/sites/default/files/blog/file/01_2_introduccion_ii.pdf). The fixture proves the complete transformation and publication path, but it is not a replacement for either the 10,000-row sample or the full dump.
+The BZip2 sources were streamed, parsed as N-Triples, and written as Zstandard-compressed Parquet. The subjects source contains 865 external BnF IRIs split immediately before the closing `>` and continued on the next physical line. Those identifiable pairs were rejoined without altering their text; the first physical line remains in `source_line`.
 
-```powershell
-bne-pipeline run --config configs/fixture.json
-```
+All 260,288,924 normalized rows passed the graph-accounting and source-line checks. No source statements remain invalid. SHA-256 hashes for every Parquet shard are provided in `artifacts/checksums.json`. Additional machine-readable information is available in:
 
-## Quality and provenance
+- `artifacts/normalization.json`
+- `artifacts/profile.json`
+- `artifacts/provenance.json`
+- `artifacts/quality.json`
+- `artifacts/schema.json`
 
-The analysis stage writes the same publication artefacts used by the other projects in this repository:
+## Intended uses and limitations
 
-- `artifacts/profile.json`: rows, distinct graph terms, object kinds, languages, and common predicates.
-- `artifacts/normalization.json`: parsed, repaired, rejected, and output counts from normalization.
-- `artifacts/schema.json`: physical schema and field semantics.
-- `artifacts/quality.json`: non-empty tables, graph accounting, normalization row agreement, and invalid-line checks.
-- `artifacts/provenance.json`: source, licence, snapshot, files, and transformations.
-- `artifacts/checksums.json`: SHA-256 and size of every published Parquet shard.
+The data is suitable for graph analysis, bibliographic research, retrieval, entity linking, authority reconciliation, and recommendation research. Consumers should create task-specific and leakage-aware evaluation splits when needed.
 
-Raw BZip2 dumps, partial downloads, local environments, and caches are excluded from Git and Hugging Face staging.
+The snapshot reflects the BNE catalogue as published in September 2023 and may contain cataloguing conventions, historical descriptions, or source errors. A relationship in the graph should not be interpreted as a current factual claim without checking the underlying catalogue record. The dataset contains metadata and identifiers, not digitized works or their full text.
 
-## Tests
+## Attribution
 
-```powershell
-.\.venv\Scripts\python.exe -m pytest
-.\.venv\Scripts\ruff.exe check .
-```
+When using the data, attribute the Biblioteca Nacional de España as the original publisher and retain the CC0 and provenance information supplied with the release.