metadata

dataset_info:
  - config_name: dav_swa
    features:
      - name: id
        dtype: int64
      - name: translation
        dtype:
          translation:
            languages:
              - dav
              - swa
    splits:
      - name: train
        num_bytes: 1578920.3838421723
        num_examples: 21329
      - name: test
        num_bytes: 394785.6161578276
        num_examples: 5333
    download_size: 1455916
    dataset_size: 1973706
  - config_name: kln_swa
    features:
      - name: id
        dtype: int64
      - name: translation
        dtype:
          translation:
            languages:
              - kln
              - swa
    splits:
      - name: train
        num_bytes: 2830217.170467162
        num_examples: 28101
      - name: test
        num_bytes: 707629.829532838
        num_examples: 7026
    download_size: 2556732
    dataset_size: 3537847
  - config_name: luo_swa
    features:
      - name: id
        dtype: int64
      - name: translation
        dtype:
          translation:
            languages:
              - luo
              - swa
    splits:
      - name: train
        num_bytes: 3510010.5175378737
        num_examples: 23446
      - name: test
        num_bytes: 877577.4824621264
        num_examples: 5862
    download_size: 3058596
    dataset_size: 4387588
configs:
  - config_name: dav_swa
    data_files:
      - split: train
        path: dav_swa/train-*
      - split: test
        path: dav_swa/test-*
  - config_name: kln_swa
    data_files:
      - split: train
        path: kln_swa/train-*
      - split: test
        path: kln_swa/test-*
  - config_name: luo_swa
    data_files:
      - split: train
        path: luo_swa/train-*
      - split: test
        path: luo_swa/test-*
license: cc-by-4.0
task_categories:
  - translation
language:
  - sw

Low-Resource Language Data: Parallel Corpora for Kiswahili and Kidaw'ida, Kalenjin, and Dholuo

Description

This dataset consists of three parallel corpora:

Kidaw'ida (Dawida)-Kiswahili (dav_swa)
Kalenjin-Kiswahili (kln_swa)
Dholuo-Kiswahili (luo_swa)

Each corpus contains approximately 30,000 sentence pairs. The dataset was created for use in training machine translation models, enabling translation from Kiswahili (the national language of Kenya) into indigenous languages.

Purpose

The primary purpose of this dataset is to facilitate the development of machine translation models for three indigenous Kenyan languages:

Kidaw'ida (Dawida)
Kalenjin
Dholuo

By providing parallel corpora with Kiswahili, this dataset aims to bridge the gap between the national language and these indigenous languages, promoting linguistic diversity and accessibility.

Dataset Details

Format: Parallel corpora (sentence pairs)
Languages: Kiswahili (swa), Kidaw'ida (dav), Kalenjin (kln), Dholuo (luo)
License: CC-BY-4.0
Task: Translation

Corpus Statistics

Kidaw'ida-Kiswahili (dav_swa):
- Train set: 21,329 examples
- Test set: 5,333 examples
- Total size: 1,973,706 bytes
Kalenjin-Kiswahili (kln_swa):
- Train set: 28,101 examples
- Test set: 7,026 examples
- Total size: 3,537,847 bytes
Dholuo-Kiswahili (luo_swa):
- Train set: 23,446 examples
- Test set: 5,862 examples
- Total size: 4,387,588 bytes

How to Use

To use this dataset for machine translation tasks:

Load the dataset using the Hugging Face Datasets library:

from datasets import load_dataset

# Load a specific language pair
dav_swa = load_dataset("kenyan-low-resource-language-data", "dav_swa")
kln_swa = load_dataset("kenyan-low-resource-language-data", "kln_swa")
luo_swa = load_dataset("kenyan-low-resource-language-data", "luo_swa")

Access the train and test splits:

train_data = dav_swa["train"]
test_data = dav_swa["test"]

Iterate through the examples:

for example in train_data:
    kidawida_text = example["translation"]["dav"]
    kiswahili_text = example["translation"]["swa"]
    # Process the text as needed

Use the data to train your machine translation model or for other NLP tasks.

Citation

If you use this dataset in your research or project, please cite it as follows:

@dataset{mbogho_2024_low_resource_language_data,
  author       = {Mbogho, Audrey and
                  Kipkebut, Andrew and
                  Wanzare, Lilian and
                  Awuor, Quin and
                  Oloo, Vivian and
                  Lugano, Rose},
  title        = {{Low-Resource Language Data: Parallel Corpora for 
                   Kiswahili and Kidaw'ida, Kalenjin, and Dholuo}},
  year         = 2024,
  publisher    = {Tech Innovators Network (THiNK) on Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/thinkKenya/kenyan-low-resource-language-data}}
}

Contributors

Creators

Audrey Mbogho (Project Manager) - United States International University Africa
Andrew Kipkebut (Data Curator) - Kabarak University
Lilian Wanzare (Data Curator) - Maseno University
Quin Awuor (Data Curator) - United States International University Africa
Vivian Oloo (Data Curator) - Maseno University
Rose Lugano (Data Curator) - University of Florida

Data Collectors

Esther Mkawanyika Nkrumah
Shalet Doreen Mkamzungu
Patience Chao Mwangola
David Mbela Mwakaba

Funding

This dataset was collected with funding from Lacuna Fund.

Updates and Future Releases

This dataset is also available on GitHub, where it will continue to be expanded and improved. Future releases will be uploaded to Hugging Face and Zenodo as new versions become available.

Contact

For questions or more information about this dataset, please contact:

Principal Investigator: Audrey Mbogho, United States International University - Africa

Acknowledgments

We would like to thank all the contributors, data collectors, and the Lacuna Fund for making this dataset possible. Their efforts contribute significantly to the preservation and technological advancement of low-resource languages in Kenya.