amanuelbyte/en-ha-translation-dataset

English-Hausa Translation Dataset

This is a merged parallel corpus for English-Hausa machine translation, created from multiple publicly available datasets. The dataset has been filtered for quality using Quality Estimation (QE) scores and deduplicated to ensure high-quality training data.

Updated: 2025-09-21

Dataset Overview

• Language Pair: English (en) to Hausa (ha)
• Total Samples After Deduplication: 156,679
• Total Samples Before Deduplication: 223,173
• Minimum QE Score: 0.6
• Creation Date: 2025-09-21

Data Sources

The dataset is compiled from the following sources, all filtered with QE >= 0.6:

Dataset Name	Rows Included	QE Score
filtered-GNOME	110	0.6625
filtered-Pontoon-Translations	1,153	0.7239
filtered-quran_multilingual_parallel	3,812	0.7094
filtered-mafand	4,446	0.7624
filtered-Tatoeba	183	0.7781
filtered-TED2020	21	0.7699
filtered-mafand-dev	971	0.7816
filtered-KDE4	6	0.7519
filtered-quran	3,731	0.7085
filtered-tico-19	2,133	0.7193
filtered-wikimedia	121,220	0.7426
filtered-AfriDocMT-health	5,901	0.7415
filtered-ntrex	1,200	0.7230
filtered-Weblate-Translations	54	0.7437
filtered-flores	794	0.7353
filtered-mafand-test	1,195	0.7900
filtered-HausaVG	7,918	0.7082
filtered-Tanzil	63,367	0.7191
filtered-smol	548	0.7119
filtered-polynews-parallel	4,410	0.7616

Processing Steps

1. Initial Filtering: Datasets were filtered using various quality criteria including language detection and semantic similarity.
2. Quality Estimation: Applied QE scoring to filter out low-quality translations (minimum score: 0.6).
3. Merging: Combined all qualifying datasets into a single corpus.
4. Deduplication:
   • Dropped identical src==tgt rows
   • Removed duplicate (src, tgt) pairs
   • Dropped duplicate sources
   • Dropped duplicate targets (removed 66,494 rows)

Dataset Structure

Each sample contains:

• src: English source text
• tgt: Hausa target text

Intended Use

This dataset is intended for training and fine-tuning machine translation models for English-Hausa translation. It can also be used for evaluation and research in low-resource language translation.

Citation

If you use this dataset in your research, please cite the original sources and mention this merged dataset.

License

Please check the licenses of the individual source datasets used in this compilation.