README.md

metadata

language:
  - kbd
  - deu
  - eng
  - fra
  - por
  - rus
  - spa
  - tur
license: cc-by-2.0
task_categories:
  - translation
pretty_name: Filtered Multilingual-to-Kabardian Tatoeba Translations Dataset
size_categories:
  - 1M<n<10M
tags:
  - translation
  - multilingual
  - parallel-corpus
  - kabardian
  - low-resource
dataset_info:
  features:
    - name: source
      dtype: string
    - name: source_lang
      dtype: string
    - name: target
      dtype: string
    - name: target_lang
      dtype: string
    - name: model_consensus_count
      dtype: int64
  splits:
    - name: train
      num_bytes: 162262527
      num_examples: 1410899
  download_size: 40616451
  dataset_size: 162262527
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Tatoeba Kabardian Filtered Dataset

Dataset Description

This dataset is a filtered version of the Multilingual-to-Kabardian Tatoeba Translations Dataset, containing higher-quality parallel sentence translations to Kabardian language (kbd). The dataset has been further filtered to ensure greater translation accuracy and consistency.

The source languages are:

Language Code	Language Name	Number of Examples
eng_Latn	English	468,894
rus_Cyrl	Russian	284,256
fra_Latn	French	177,888
deu_Latn	German	165,468
por_Latn	Portuguese	121,100
spa_Latn	Spanish	107,046
tur_Latn	Turkish	86,247

Dataset Summary

The dataset consists of high-quality parallel translations from multiple languages to Kabardian, containing approximately 1.4M examples. This filtered version applies stricter quality controls compared to the original dataset, making it particularly valuable for tasks requiring high-precision translations. It serves as an important resource for low-resource language translation research and applications involving the Kabardian language.

Languages

Source languages:

• German (deu_Latn)
• English (eng_Latn)
• French (fra_Latn)
• Portuguese (por_Latn)
• Russian (rus_Cyrl)
• Spanish (spa_Latn)
• Turkish (tur_Latn)

Target language:

• Kabardian (kbd_Cyrl)

Dataset Structure

Data Instances

Example of a data instance:

{
  "source": "He's always happy.",
  "source_lang": "eng_Latn",
  "target": "Ар сыт щыгъуи насыпыфIэщ.",
  "target_lang": "kbd_Cyrl",
  "model_consensus_count": 3
}

Data Fields

• source: Original text in source language
• source_lang: Language code of the source text (with script identifier, e.g., "eng_Latn")
• target: Translated text in Kabardian
• target_lang: Language code for Kabardian (kbd_Cyrl)
• model_consensus_count: Count of model versions that produced this identical translation. Higher values indicate greater consensus among different model versions, serving as a reliability metric for the translation quality

Dataset Creation

Source Data

The source data comes from Tatoeba, a free collaborative online database of example sentences, with translations processed through various versions of NLLB-200 models.

Initial Data Collection and Normalization

1. Sentences were extracted from Tatoeba's database
2. Filtered by language pairs
3. Translated using various versions of NLLB-200-kbd model
4. Translation quality was assessed using panagoa/LaBSE-kbd-v0.1 model
5. Additional filtering was applied based on a consensus mechanism - only translations that were consistently produced by multiple model versions (tracked in model_consensus_count field) were included in this dataset

Annotations

The translations were generated using different versions of NLLB-200-kbd model. This filtered dataset represents translations that met more stringent quality criteria than the original dataset, particularly focusing on consensus among different model versions.

Considerations for Using the Data

Social Impact of Dataset

This dataset contributes to the linguistic resources available for the Kabardian language, which is considered a low-resource language in NLP. It can help in developing machine translation systems and other language technologies for Kabardian speakers, potentially contributing to language preservation and accessibility.

Discussion of Biases

As with many translation datasets, there may be biases in the sentence selection and translation. Users should be aware that:

• The dataset may not represent the full diversity of contexts, dialects, and expressions in Kabardian language
• Machine-translated content may contain systemic biases present in the original translation models
• The filtering process, while improving quality, may introduce selection biases

Additional Information

Dataset Curators

This dataset was curated by [panagoa] as a refined version of the original Tatoeba Kabardian dataset.

Licensing Information

This dataset is licensed under Creative Commons Attribution 2.0.

Citation Information

@misc{tatoeba_kbd_filtered_2025,
    title={Filtered Multilingual-to-Kabardian Tatoeba Translations Dataset},
    author={Adam Panagov},
    year={2025},
    publisher={Hugging Face},
    howpublished={\url{https://huggingface.co/datasets/panagoa/tatoeba_kbd_filtered}}
}

Contributions

Thanks to @Tatoeba for providing the source data and to the Kabardian language community.

Usage Examples

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("panagoa/tatoeba_kbd_filtered")

# Example usage
for example in dataset['train'].select(range(5)):
    print(f"Source ({example['source_lang']}): {example['source']}")
    print(f"Target ({example['target_lang']}): {example['target']}")
    print(f"Model consensus count: {example['model_consensus_count']}")
    print("---")