sotalab/en-vi-ja-200k-triplets

EN-VI-JA 200K Triplets (Commercial-Friendly)

A high-quality parallel corpus of 199,770 English-Vietnamese-Japanese sentence triplets for multilingual machine translation, built exclusively from commercial-use friendly datasets.

Key Features

• Commercial Use Allowed: Licensed under CC-BY-SA 4.0
• High Quality: 10-step quality pipeline with language validation, deduplication, and coherence filtering
• Trilingual Alignment: Each row contains the same sentence in EN, VI, and JA

Dataset Statistics

Split	Examples
Train	179,793
Validation	9,988
Test	9,989
Total	199,770

Source Distribution

Source	Triplets	Percentage	License
CCAligned	143,320	71.7%	Permissive
WikiMatrix	26,780	13.4%	CC-BY-SA
HPLT v2	19,842	9.9%	CC0
KDE4	8,398	4.2%	FOSS
Tatoeba	1,430	0.7%	CC-BY 2.0

Quality Metrics

• Average Quality Score: 0.586
• Score Distribution: 74% in 0.4-0.6 range, 26% in 0.6-0.8 range

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("sotalab/en-vi-ja-200k-triplets")

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

# Example usage
for example in train.select(range(3)):
    print(f"EN: {example['en']}")
    print(f"VI: {example['vi']}")
    print(f"JA: {example['ja']}")
    print()

Data Format

Field	Type	Description
en	string	English sentence
vi	string	Vietnamese translation
ja	string	Japanese translation
source	string	Origin dataset (CCAligned, WikiMatrix, etc.)
quality_score	float	Quality score (0.0-1.0)

Processing Pipeline

1. Download: Fetched parallel corpora from OPUS (HPLT, CCAligned, WikiMatrix, Tatoeba, KDE4)
2. Cleaning: Unicode normalization (NFKC), HTML/URL removal
3. Length Filtering: 5-100 tokens, length ratio ≤ 2.5
4. Language Validation: FastText language ID + script validation (Vietnamese diacritics, Japanese kana/kanji)
5. Content Filtering: Removed boilerplate, code fragments, low-quality text
6. Triplet Creation: Pivot join on normalized English text
7. Coherence Validation: Cross-lingual number/punctuation consistency
8. Deduplication: Polars-based near-duplicate removal
9. Quality Scoring: Multi-factor scoring (language, length, source quality)
10. Selection: Top 200K by quality score

License

CC-BY-SA 4.0 - Commercial use is permitted with attribution.

See LICENSE for full terms and source attribution requirements.

Citations

If you use this dataset, please cite the source corpora:

@inproceedings{elkishky2020ccaligned,
  title={CCAligned: A Massive Collection of Cross-lingual Web-Document Pairs},
  author={El-Kishky, Ahmed and Chaudhary, Vishrav and Guzman, Francisco and Koehn, Philipp},
  booktitle={EMNLP},
  year={2020}
}

@inproceedings{schwenk2021wikimatrix,
  title={WikiMatrix: Mining 135M Parallel Sentences in 1620 Language Pairs from Wikipedia},
  author={Schwenk, Holger and Chaudhary, Vishrav and Sun, Shuo and Gong, Hongyu and Guzman, Francisco},
  booktitle={EACL},
  year={2021}
}

@inproceedings{hplt2024,
  title={A New Massive Multilingual Dataset for High-Performance Language Technologies},
  author={HPLT Consortium},
  booktitle={LREC-COLING},
  year={2024}
}

Use Cases

• Fine-tuning NLLB-200 for EN-VI-JA translation
• Training multilingual translation models
• Pivot-based VI↔JA translation research
• Cross-lingual NLP experiments

Limitations

• Web-crawled data may contain noise or occasional mistranslations
• Some triplets may have minor alignment errors
• Content reflects web sources (news, Wikipedia, software localization, etc.)

Contact

For questions or issues, please open an issue on this repository.