belumind/en-vi-ja-curated-27k-triplets

EN-VI-JA Curated 27K Triplets

A high-quality, commercially-friendly English-Vietnamese-Japanese parallel corpus with embedding-verified alignment.

Dataset Summary

Split	Count
Train	24,750
Validation	1,375
Test	1,376
Total	27,501

Quality Metrics

This dataset has been rigorously filtered using LaBSE (Language-agnostic BERT Sentence Embeddings) for cross-lingual alignment verification.

Metric	Score
Average Quality Score	0.7648
EN-VI Similarity	0.8653
EN-JA Similarity	0.8633
VI-JA Similarity	0.8396

All triplets have minimum similarity >= 0.70 across all language pairs.

Quality Pipeline

1. Script Validation: Verified correct scripts (Latin for English, Vietnamese diacritics, Japanese Hiragana/Katakana/Kanji)
2. Length Ratio Filtering: Removed misaligned pairs with abnormal length ratios
3. Embedding Similarity: LaBSE embeddings with threshold >= 0.70 for all pairs
4. Semantic Deduplication: Removed near-duplicates using embedding clustering
5. Quality Score Ranking: Ranked by combined trilingual similarity

Data Sources (All Commercial-Friendly)

Source	License
Tatoeba	CC-BY 2.0
WikiMatrix	CC-BY-SA
KFTT	CC-BY-SA 3.0
LaboroParaCorpus	CC0 (Public Domain)
AMIMeeting	CC-BY 4.0
KDE4	FOSS (permissive)
Ubuntu	FOSS (permissive)

License

CC-BY-SA (Creative Commons Attribution-ShareAlike)

This dataset is safe for commercial use with proper attribution.

Usage

from datasets import load_dataset

dataset = load_dataset("sotalab/en-vi-ja-curated-27k-triplets")

# Access splits
train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

# Example
print(train[0])
# {'en': '...', 'vi': '...', 'ja': '...'}

Data Format

Each example contains:

• en: English sentence
• vi: Vietnamese translation
• ja: Japanese translation

Citation

If you use this dataset, please cite:

@dataset{en_vi_ja_curated_triplets,
  title={EN-VI-JA Curated 27K Triplets},
  author={SotaLab},
  year={2024},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/sotalab/en-vi-ja-curated-27k-triplets}
}

Acknowledgments

This dataset was created from the following projects:

• OPUS Project - Open parallel corpus collection
• Tatoeba - Collaborative sentence collection
• WikiMatrix - Wikipedia parallel sentences
• KFTT - Kyoto Free Translation Task
• LaboroParaCorpus - Laboro.AI parallel corpus
• sentence-transformers/LaBSE - Language-agnostic embeddings