README.md

---
language:
- es
license: cc-by-sa-4.0
multilinguality: monolingual
task_categories:
- text-classification
task_ids: []
tags:
- mteb
- clustering
- wikipedia
- spanish
- embeddings
- evaluation
annotations_creators:
- derived
source_datasets:
- wikimedia/wikipedia
pretty_name: SpanishWikiClustering
size_categories:
- 10K<n<100K
---

# SpanishWikiClustering

Clustering benchmark for Spanish embedding models, designed for [MTEB](https://github.com/embeddings-benchmark/mteb) as **SpanishWikiClusteringP2P**.

Given opening paragraphs from Spanish Wikipedia articles, models must group them by topic without seeing the labels. This task format is inspired by [WikiClusteringP2P](https://github.com/jhrystrom/wiki-clustering) (Rystrøm, 2024), which covers 14 languages — but not Spanish. This dataset fills that gap.

## Dataset

- **Source**: Spanish Wikipedia via CirrusSearch dump (December 2025)
- **Format**: MTEB clustering — `sentences` (list of texts) + `labels` (list of categories)
- **Samples**: 30 independent clustering tasks
- **Articles per sample**: 900 (100 per category, stratified)
- **Split**: test

### Categories (9)

| Category | Pool size |
|----------|-----------|
| Psicología | 768K articles |
| Lenguaje | 210K articles |
| Ciencias_naturales | 210K articles |
| Geografía | 176K articles |
| Historia | 63K articles |
| Organizaciones | 23K articles |
| Literatura | 21K articles |
| Deporte | 12K articles |
| Fenómenos_naturales | 6K articles |

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("ClementeH/SpanishWikiClustering", split="test")
# 30 rows, each with 900 sentences and 900 labels

sample = dataset[0]
print(len(sample["sentences"]))  # 900
print(set(sample["labels"]))     # 9 categories
```

### Evaluate with MTEB

```python
import mteb

model = mteb.get_model("intfloat/multilingual-e5-base")
task = mteb.get_task("SpanishWikiClusteringP2P")
results = mteb.evaluate(model, tasks=[task])
```

## Benchmark Results

Evaluated with MTEB's `v_measure` metric (higher is better). 30 samples of 900 articles each.

| Model | v_measure |
|-------|-----------|
| intfloat/multilingual-e5-large-instruct | **0.3679** |
| BAAI/bge-m3 | 0.3308 |
| sentence-transformers/paraphrase-multilingual-mpnet-base-v2 | 0.3296 |
| intfloat/multilingual-e5-large | 0.3249 |
| intfloat/multilingual-e5-base | 0.3226 |
| intfloat/multilingual-e5-small | 0.3194 |
| nomic-ai/nomic-embed-text-v1.5 | 0.3183 |
| sentence-transformers/distiluse-base-multilingual-cased-v2 | 0.2999 |
| sentence-transformers/stsb-xlm-r-multilingual | 0.2904 |
| Snowflake/snowflake-arctic-embed-m-v1.5 | 0.2883 |
| sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2 | 0.2863 |
| sentence-transformers/LaBSE | 0.2861 |
| minishlab/potion-multilingual-128M | 0.2843 |
| sentence-transformers/static-similarity-mrl-multilingual-v1 | 0.2527 |

For reference, `WikiClusteringP2P.v2` (e5-base) scores for comparable languages: Czech 0.3198, Albanian 0.2936, Danish 0.1920.

## Replication

The pipeline to regenerate the dataset from scratch is at [Clemente-H/clustering_embeddings_es](https://github.com/Clemente-H/clustering_embeddings_es), adapted from [jhrystrom/wiki-clustering](https://github.com/jhrystrom/wiki-clustering) to support Spanish Wikipedia's MediaWiki 1.42+ schema.

```bash
git clone https://github.com/Clemente-H/clustering_embeddings_es
cd clustering_embeddings_es
pip install -r requirements.txt

# Full pipeline (~10 GB download, ~20 GB RAM)
python run.py

# Skip download if dumps are already in local_data/
python run.py --skip-download --n-articles 900 --n-turns 30
```

### Key adaptations vs. original wiki-clustering

The original pipeline targets MediaWiki ≤1.41. Spanish Wikipedia runs MW 1.42+, which introduced breaking changes:

1. **`categorylinks` schema change**: `cl_to` (category name string) was replaced by `cl_target_id` (foreign key to the new `linktarget` table)
2. **SQL encoding**: category names in SQL dumps are stored as latin1-misinterpreted UTF-8 — matching requires re-encoding before joining
3. **CirrusSearch over XML**: the XML dump requires ~64 GB RAM; CirrusSearch is NDJSON, streamable with ~8 GB
4. **Category curation**: Spanish Wikipedia's automatic top-level categories are too coarse (5 nodes). We use 9 thematic categories from the 2nd level, curated to avoid nesting, inflated person-profession categories, and semantic ambiguity
5. **Two-list config**: `curated_categories` anchors the tree walk (prevents pool contamination from excluded categories); `sample_categories` is the subset used for labeling. Excluded categories must remain as anchors so their articles are not reassigned to neighboring pools

## Citation

If you use this dataset, please cite:

```bibtex
@misc{henriquez2026spanishwikiclustering,
  author       = {Henriquez, Clemente},
  title        = {{SpanishWikiClustering}: A Wikipedia-based Clustering Benchmark for Spanish},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/ClementeH/SpanishWikiClustering}},
}
```

If you use the replication pipeline, please also cite the original wiki-clustering work it is based on:

```bibtex
@misc{rystrom2024wikiclustering,
  author       = {Rystr{\o}m, Jonathan},
  title        = {Wiki Clustering},
  year         = {2024},
  howpublished = {\url{https://github.com/jhrystrom/wiki-clustering}},
}
```

If you use this dataset via MTEB, please also cite:

```bibtex
@article{muennighoff2022mteb,
  author  = {Muennighoff, Niklas and Tazi, Nouamane and Magne, Lo{\"i}c and Reimers, Nils},
  title   = {{MTEB}: Massive Text Embedding Benchmark},
  journal = {arXiv preprint arXiv:2210.07316},
  year    = {2022},
  url     = {https://arxiv.org/abs/2210.07316},
}

@article{enevoldsen2025mmteb,
  title   = {{MMTEB}: Massive Multilingual Text Embedding Benchmark},
  author  = {Enevoldsen, Kenneth and others},
  journal = {arXiv preprint arXiv:2502.13595},
  year    = {2025},
  url     = {https://arxiv.org/abs/2502.13595},
}
```