SpanishWikinewsClustering

SpanishWikinewsClustering is a small, reproducible Spanish news clustering dataset for MTEB. It has two configurations sharing exactly the same 800 source articles and eight thematic labels: s2s evaluates titles; p2p evaluates each title followed by a cleaned lead.

Dataset summary

Configuration	Text	Rows	Labels	Split
s2s	Title	800	8	test
p2p	Title + cleaned lead (maximum 2,000 characters)	800	8	test

The labels are Arte, cultura y entretenimiento; Ciencia y tecnología; Deportes; Desastres y accidentes; Economía y Negocios; Judicial; Política; and Sociedad. Each has exactly 100 rows. Files are flat JSONL with the fields sentences and labels, ready for MTEB's current AbsTaskClustering implementation.

Source, license, and attribution

The source is the Spanish Wikinews Wikimedia dump in the 20260801 dump directory. The SHA-256 of eswikinews-20260801-pages-articles.xml.bz2 is bfb32ff73532cfffecf5db13e3712409ad9134459ea1d1927a8f93429be87109.

The source text is licensed under CC BY 2.5. This dataset is a derived work. Please attribute the original content to contributors to Spanish Wikinews and link to the Wikimedia dump. The cleaning, filtering, and sampling described below are additional modifications.

Construction and quality controls

The build pipeline uses the same snapshot's article, page, categorylinks, and linktarget dumps. It keeps main-namespace non-redirect pages with exactly one thematic category after transitive category-graph propagation. It excludes 549 multi-item pages containing “Noticias breves”, removes exact and lexical near-duplicates (five-token-shingle Jaccard threshold 0.85), and samples without replacement with seed 42.

Author manual review was stratified by label (20 items × 8). In its last review round, 155 of 160 sampled items were marked pass; one ambiguous and four wrong-label cases were removed. Those five selected items were replaced deterministically within their original labels, while the 155 approved items were retained. Earlier review rounds produced 45 conservative global title exclusions. LLM-assisted spot QA (Claude and Mistral, 16 items) was used only to surface items for human inspection; it was not used as a source of labels.

This is author manual review + LLM-assisted spot QA, not independent human validation.

Limitations

• The source is a small, historical, community-edited news corpus; it is not representative of all Spanish varieties, contemporary news, or every topic.
• Labels derive from Wikinews category structure and are intentionally strict/mono-label.
• The dataset is intended for comparative embedding evaluation, not topic-classification training or factual inference.

Citation

@misc{henriquez2026spanishwikinewsclustering,
  author = {Henríquez, Clemente},
  title = {SpanishWikinewsClustering},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/ClementeH/SpanishWikinewsClustering}
}

Reproducibility

The extraction and auditing pipeline is maintained in the ES-MTEB project: evaluations/wikinews/build_dataset.py. The accompanying report.json records input hashes, pool counts, filtering counts, parameters, and post-audit selected-item replacements.