---
dataset_info:
  features:
  - name: en
    dtype: string
  - name: yo
    dtype: string
  splits:
  - name: train
    num_bytes: 6765615
    num_examples: 22626
  download_size: 4149995
  dataset_size: 6765615
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
# Yoruba-English Translation Dataset

A curated parallel corpus for machine translation between English (en) and Yoruba (yo), designed for low-resource machine translation research and applications.

## 📊 Dataset Overview

- **Language Pair**: English → Yoruba (en-yo)
- **Total Sentence Pairs**: 20,703
- **Quality Threshold**: QE ≥ 0.6
- **Processing Date**: September 22, 2025
- **Repository**: [amanuelbyte/yo-en-translation-dataset](https://huggingface.co/amanuelbyte/yo-en-translation-dataset)

## 🎯 Key Features

- **Low-Resource Optimized**: Tailored quality thresholds for Yoruba language characteristics
- **Deduplicated**: Extensive deduplication process removes redundant content
- **Multi-Domain**: Covers diverse domains including technical, religious, educational, and general content
- **Quality Assured**: All sentence pairs meet minimum Quality Estimation score of 0.6

## 📈 Dataset Statistics

### Data Volume
- **Before Deduplication**: 32,796 sentence pairs
- **After Deduplication**: 20,703 sentence pairs
- **Deduplication Ratio**: 36.9% reduction

### Quality Metrics
- **Average QE Score**: 0.68
- **Highest Quality Dataset**: Tatoeba (QE: 0.7137)
- **Lowest Quality Dataset**: Ubuntu (QE: 0.6452)

## 📚 Data Sources

This dataset combines **20 high-quality sources**:

| Source | Dataset | Pairs | QE Score | Domain |
|--------|---------|-------|----------|---------|
| wikimedia | Wikipedia Articles | 4,773 | 0.7035 | General |
| mafand | Mafand Dataset | 3,985 | 0.6951 | General |
| mafand-dev | Mafand Development | 3,985 | 0.6951 | General |
| mafand-test | Mafand Test Set | 3,985 | 0.6951 | General |
| AfriDocMT-health | African Medical Documents | 3,701 | 0.6721 | Medical |
| menyo20k_mt-test | Menyo20k Test Set | 3,682 | 0.6947 | General |
| menyo20k_mt-train | Menyo20k Training | 4,585 | 0.6997 | General |
| menyo20k_mt-dev | Menyo20k Development | 1,432 | 0.6853 | General |
| quran_multilingual_parallel | Multilingual Quran | 1,002 | 0.6568 | Religious |
| ntrex | News Articles | 602 | 0.6780 | News |
| Pontoon-Translations | Mozilla Translations | 553 | 0.6795 | Technical |
| smol | Smol Dataset | 153 | 0.6694 | General |
| GlobalVoices | Global Voices | 61 | 0.6769 | News |
| opus-100 | OPUS-100 Corpus | 59 | 0.6557 | General |
| opus-100-dev | OPUS-100 Development | 59 | 0.6557 | General |
| opus-100-test | OPUS-100 Test Set | 59 | 0.6557 | General |
| GNOME | Software Localization | 59 | 0.6557 | Technical |
| Tatoeba | Community Translations | 21 | 0.7137 | General |
| Weblate-Translations | Weblate Platform | 33 | 0.7056 | General |
| Ubuntu | Ubuntu Localization | 7 | 0.6452 | Technical |

## 🛠️ Processing Pipeline

The dataset underwent comprehensive processing optimized for Yoruba:

1. **Rule-based Filtering**: Length constraints, character validation
2. **Semantic Filtering**: Sentence embedding similarity (threshold: 0.7)
3. **Language Detection**: FastText + AfroLID models for accurate language identification
4. **Quality Estimation**: Africomet QE model for translation quality scoring
5. **Sentence-level QE Filtering**: Only pairs with QE ≥ 0.6 retained (optimized for Yoruba)
6. **Deduplication**: 4-stage process removing identical and similar content

### Deduplication Process
1. Remove identical source-target pairs
2. Remove duplicate (source, target) combinations
3. Remove duplicate source sentences
4. Remove duplicate target sentences

## 💾 Dataset Format

The dataset is available in Hugging Face `datasets` format with the following structure:

```python
from datasets import load_dataset

dataset = load_dataset("amanuelbyte/yo-en-translation-dataset")
print(dataset)

# Dataset structure:
# DatasetDict({
#     "train": Dataset({
#         features: ['en', 'yo'],
#         num_rows: 20703
#     })
# })
```

### Columns
- `en`: English source text
- `yo`: Yoruba target text

## 📖 Usage Examples

### Loading the Dataset
```python
from datasets import load_dataset

# Load the dataset
dataset = load_dataset("amanuelbyte/yo-en-translation-dataset")

# Access the training split
train_data = dataset["train"]

# Get a sample
sample = train_data[0]
print(f"English: {sample['en']}")
print(f"Yoruba: {sample['yo']}")
```

### For Machine Translation Training
```python
# Prepare data for training
source_texts = train_data["en"]
target_texts = train_data["yo"]

# Statistics
print(f"Dataset size: {len(train_data)}")
print(f"Average source length: {sum(len(text.split()) for text in source_texts) / len(source_texts)".1f"}")
print(f"Average target length: {sum(len(text.split()) for text in target_texts) / len(target_texts)".1f"}")
```

## 🎓 Applications

This dataset is particularly suitable for:
- **Low-Resource MT**: Training models for under-resourced Yoruba language
- **Cross-lingual Transfer**: Multilingual NLP tasks involving African languages
- **Language Modeling**: Pretraining language models for Yoruba
- **Quality Estimation Research**: Studies on QE thresholds for low-resource languages
- **Linguistic Studies**: Analysis of Yoruba-English language characteristics

## 🔍 Quality Assessment

### Dataset-Specific Insights
- **Optimized QE Threshold**: 0.6 (vs 0.7 for higher-resource languages)
- **Quality Distribution**: Reflects the challenges of Yoruba language processing
- **Source Diversity**: Combines multiple domains for robust training data
- **Deduplication Effectiveness**: 36.9% reduction indicates significant data cleaning

### Quality Metrics Summary
- **Top Performing Sources**: Tatoeba, Weblate-Translations, Wikimedia
- **Largest Contributors**: Mafand datasets (11,955 pairs total)
- **Domain Coverage**: Technical, religious, news, and general content

## 📄 License

This dataset is released under [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/) license. The individual source datasets maintain their original licenses.

## 🤝 Citation

If you use this dataset in your research, please cite:

```bibtex
@dataset{yo-en-translation-dataset-2025,
  title={Yoruba-English Translation Dataset},
  author={Amanuel Tewolde},
  year={2025},
  publisher={Hugging Face},
  journal={Hugging Face Hub},
  howpublished={\url{https://huggingface.co/amanuelbyte/yo-en-translation-dataset}}
}
```

## 📞 Contact

For questions or issues regarding this dataset, please contact:
- **Maintainer**: Amanuel Tewolde
- **Email**: amanuelbyte@protonmail.com
- **Organization**: Hugging Face Community

## 🔗 Related Resources

- [AfroLID Model](https://huggingface.co/UBC-NLP/afrolid_1.5) - Language identification
- [Africomet QE](https://huggingface.co/masakhane/africomet-qe-stl) - Quality estimation
- [Processing Pipeline](https://github.com/amanuelbyte/mt-data-processing) - Source code
- [Afrikaans Dataset](https://huggingface.co/amanuelbyte/af-en-translation-dataset) - Related work

---

**Note**: This Yoruba dataset uses optimized quality thresholds (QE ≥ 0.6) compared to higher-resource languages, reflecting the challenges and characteristics of low-resource African language processing. The lower threshold ensures sufficient training data while maintaining quality standards appropriate for Yoruba language resources.