belumind/en-vi-ja-curated-500k-triplets

EN-VI-JA Curated 500K Triplets

High-quality English-Vietnamese-Japanese translation triplets curated from OPUS parallel corpora.

Dataset Statistics

Split	Count
Train	446,252
Validation	24,792
Test	24,792
Total	495,836

Quality Metrics

• Average Quality Score: 0.794
• Average LaBSE Score: 0.859
• LaBSE Threshold: 0.6

LaBSE Score Distribution

Range	Count
0.9-1.0	177,022
0.8-0.9	217,463
0.7-0.8	76,727
0.5-0.7	24,624

Source Distribution

Source	Count	Percentage
CCAligned	370,871	74.8%
HPLT	98,412	19.8%
WikiMatrix	13,887	2.8%
KDE4	8,364	1.7%
LaboroParaCorpus	1,934	0.4%
Tatoeba	1,178	0.2%
Ubuntu	1,159	0.2%
GNOME	11	<0.1%
AMIMeeting	11	<0.1%
KFTT	9	<0.1%

Quality Pipeline (19 Steps)

V2 Steps (10)

1. Language validation (FastText + script)
2. Deduplication
3. Cross-lingual coherence
4. Text normalization
5. Content quality filters
6. Japanese tokenization
7. Quality-based selection
8. Source quality weighting
9. Triplet length ratio
10. Encoding handling

V3 Steps (9)

11. LaBSE semantic alignment
12. Broken English detection
13. Web artifact removal
14. B2B spam filtering
15. NSFW content filtering
16. Fragment detection
17. MT artifact detection
18. Script ratio validation
19. Character ratio validation

Quality Thresholds

Parameter	Value
FastText confidence	0.65
Vietnamese diacritic ratio	0.02
Japanese script ratio	0.10
LaBSE threshold	0.60
Min tokens	3
Max tokens	100
Max length ratio	2.5

Usage

from datasets import load_dataset

dataset = load_dataset("sotalab/en-vi-ja-curated-500k-triplets")

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

# Example
print(train[0])
# {'en': '...', 'vi': '...', 'ja': '...', 'source': '...', 'quality_score': 0.85, 'labse_score': 0.92}

License

This dataset is released under the CC-BY-4.0 license.

Citation

If you use this dataset, please cite:

@dataset{sotalab2024envija500k,
  title={EN-VI-JA Curated 500K Triplets},
  author={SotaLab},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/sotalab/en-vi-ja-curated-500k-triplets}
}