aimamba/WikiMatrix-en-lv

WikiMatrix EN-LV

Dataset Description

597,480 English-Latvian parallel sentence pairs mined from Wikipedia using the WikiMatrix methodology.

Method

1. Source: English and Latvian Wikipedia article dumps (April 2026)
2. Embeddings: LaBSE (Language-agnostic BERT Sentence Embeddings)
3. Retrieval: FAISS approximate nearest-neighbor search (IndexFlatIP)
4. Scoring: Margin-based scoring — margin(x, y) = cos(x, y) / [(Σ cos(x, nn_y) + Σ cos(y, nn_x)) / (2k)]
5. Filtering: Pairs with margin score ≥ 1.04 retained

Columns

Column	Type	Description
en	string	English sentence
lv	string	Latvian sentence
score	float	Margin similarity score (higher = more confident alignment)

Splits

Split	Examples
train	537,732
validation	29,874
test	29,874

Quality Thresholds

The score column can be used to filter for higher-quality pairs:

Threshold	Approx. Pairs	Quality
≥ 1.04	597,480	All pairs (full dataset)
≥ 1.06	~500,000	Recommended for training
≥ 1.10	~350,000	High confidence
≥ 1.20	~150,000	Very high confidence

from datasets import load_dataset

ds = load_dataset("aimamba/WikiMatrix-en-lv")

# Filter for high-quality pairs
high_quality = ds["train"].filter(lambda x: x["score"] >= 1.06)
print(f"High-quality pairs: {len(high_quality)}")

License

CC-BY-SA 4.0 (inherited from Wikipedia)

Citation

@misc{wikimatrix-en-lv-2026,
  title={WikiMatrix EN-LV: English-Latvian Parallel Corpus from Wikipedia},
  author={aimamba},
  year={2026},
  howpublished={\url{https://huggingface.co/datasets/aimamba/WikiMatrix-en-lv}},
  note={597,480 sentence pairs mined using LaBSE + FAISS}
}

Acknowledgments

• LaBSE — Feng et al., 2022
• WikiMatrix — Schwenk et al., 2019 (methodology inspiration)
• FAISS — Facebook AI Research