belumind/en-vi-ja-300k-triplets

EN-VI-JA 300K Triplet Dataset

A high-quality parallel corpus of 300,000 sentence-level aligned triplets across English, Vietnamese, and Japanese.

Dataset Information

• Languages: English (en), Vietnamese (vi), Japanese (ja)
• Size: 300,000 triplets
• Format: JSONL with en, vi, ja keys

Data Sources

This dataset is curated from OPUS parallel corpora:

Corpus	Language Pair	Description
CCMatrix v1	EN-VI, EN-JA	Web-crawled parallel sentences
OpenSubtitles v2024	EN-VI	Movie and TV subtitle translations
JParaCrawl v3.0	EN-JA	Japanese web parallel corpus
WikiMatrix v1	EN-JA	Wikipedia-extracted translations

Data Pipeline

1. Collection - Downloaded parallel sentence pairs from OPUS
2. Cleaning - Removed noise, special characters, and malformed text
3. Filtering - Kept sentences with appropriate length and quality
4. Language Verification - Validated correct languages in each pair
5. Deduplication - Removed duplicate and near-duplicate sentences
6. Triplet Creation - Joined EN-VI and EN-JA pairs using English as pivot
7. Final Sampling - Selected 300K high-quality triplets

Usage

from datasets import load_dataset

dataset = load_dataset("sotalab/en-vi-ja-300k-triplets")

# Access a sample
sample = dataset["train"][0]
print(f"English: {sample['en']}")
print(f"Vietnamese: {sample['vi']}")
print(f"Japanese: {sample['ja']}")

Author

Bao Bao

License

This dataset is released for research purposes only.

This dataset is derived from multiple sources with different licenses. Due to JParaCrawl's license restrictions, this dataset cannot be released under CC-BY-4.0 and is restricted to non-commercial research use only.

Source Licenses

• CCMatrix - CC-BY-4.0
• OpenSubtitles - Open license
• JParaCrawl - Research use only (non-commercial)
• WikiMatrix - CC-BY-SA-3.0

Links

• Dataset Viewer Space