README.md

metadata

dataset_info:
  features:
    - name: text_bem
      dtype: string
    - name: text_en
      dtype: string
  splits:
    - name: train
      num_bytes: 2041746
      num_examples: 20121
  download_size: 1203727
  dataset_size: 2041746
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
task_categories:
  - translation
language:
  - af
  - en

Dataset Details

This dataset contains Bemba-to-English sentences which is intended to machine translation task. This dataset was gained from Tatoeba-Translations repository by Yasmin Moslem. Tatoeba is a dataset of sentences and its translations [1].

Preprocessing Notes

There are several preprocessing processes done with this dataset.

1. Take for about ~300k English data from Tatoeba.
2. Translating that English data using our MT model and keep the translating score.
3. Keep the data that has score above 0,7.
4. Add <bt> tag in front of the Bemba sentences to indicate that the data is generated by back-translating method.

Dataset Structure

DatasetDict({
    train: Dataset({
        features: ['text_bem', 'text_en'],
        num_rows: 20121
    })
})

Reference

1. https://tatoeba.org/