README.md

---
language:
- kbd
- deu
- eng
- fra
- por
- rus
- spa
- tur
license: cc-by-2.0
task_categories:
- translation
pretty_name: Tatoeba Translations Dataset
size_categories:
- 1M<n<10M
tags:
- translation
- multilingual
- parallel-corpus
dataset_info:
  features:
  - name: source_tatoeba_id
    dtype: int64
  - name: source_lang
    dtype: string
  - name: source_text
    dtype: string
  - name: target_text
    dtype: string
  - name: similarity_score
    dtype: float64
  - name: nllb_model_version
    dtype: string
  - name: __index_level_0__
    dtype: int64
  splits:
  - name: deu
    num_bytes: 78507756
    num_examples: 530327
  - name: eng
    num_bytes: 343953362
    num_examples: 2252679
  - name: fra
    num_bytes: 72440759
    num_examples: 502436
  - name: por
    num_bytes: 49640754
    num_examples: 340764
  - name: rus
    num_bytes: 156138437
    num_examples: 907621
  - name: spa
    num_bytes: 48255668
    num_examples: 330634
  - name: tur
    num_bytes: 129012850
    num_examples: 868372
  download_size: 278574348
  dataset_size: 877949586
configs:
- config_name: default
  data_files:
  - split: deu
    path: data/deu-*
  - split: eng
    path: data/eng-*
  - split: fra
    path: data/fra-*
  - split: por
    path: data/por-*
  - split: rus
    path: data/rus-*
  - split: spa
    path: data/spa-*
  - split: tur
    path: data/tur-*
---

# Tatoeba Translations Dataset

## Dataset Description

This dataset contains parallel sentence translations from [Tatoeba](https://tatoeba.org/) to Kabardian language (kbd), filtered by similarity score. The source languages are:

- German (deu)
- English (eng)
- French (fra)
- Portuguese (por)
- Russian (rus)
- Spanish (spa)
- Turkish (tur)

All translations in this dataset are paired with Kabardian (kbd) as the target language.

### Dataset Summary

The dataset consists of high-quality parallel translations between multiple language pairs, where each translation pair has been filtered based on similarity scores to ensure quality. Each language section contains translations to/from Russian, making this dataset particularly useful for multilingual translation tasks involving Russian.

### Languages

The dataset includes translations from the following source languages to Kabardian:

Source languages:
- German (deu) [ISO 639-3](https://iso639-3.sil.org/)
- English (eng) [ISO 639-3](https://iso639-3.sil.org/)
- French (fra) [ISO 639-3](https://iso639-3.sil.org/)
- Portuguese (por) [ISO 639-3](https://iso639-3.sil.org/)
- Russian (rus) [ISO 639-3](https://iso639-3.sil.org/)
- Spanish (spa) [ISO 639-3](https://iso639-3.sil.org/)
- Turkish (tur) [ISO 639-3](https://iso639-3.sil.org/)

Target language:
- Kabardian (kbd) [ISO 639-3](https://iso639-3.sil.org/)

## Dataset Structure

### Data Instances

Each instance in the dataset contains:
```python
{
    'source_tatoeba_id': 7627627,                 # Tatoeba sentence ID
    'source_lang': 'eng',                         # Source language code
    'source_text': 'He's always happy.',          # Original text
    'target_text': 'Ар сыт щыгъуи насыпыфIэщ.',   # Translated text
    'similarity_score': 0.9283,                   # Translation similarity score
    'nllb_model_version': 'rus-kbd-v1.3.4'        # Model used for verification
}
```

### Data Fields

- `source_tatoeba_id`: Unique identifier from Tatoeba
- `source_lang`: Language code of the source text
- `source_text`: Original text in source language
- `target_text`: Translated text (predominantly in Russian)
- `similarity_score`: Similarity score between source and target texts
- `nllb_model_version`: Version of the translation model used for verification

### Data Splits

The dataset is organized by source languages, with each language paired with Kabardian as the target language. The similarity thresholds represent the 75th percentile of similarity scores for each language pair:
- deu: German-Kabardian pairs (75th percentile similarity: 0.8711)
- eng: English-Kabardian pairs (75th percentile similarity: 0.8800)
- fra: French-Kabardian pairs (75th percentile similarity: 0.8716)
- por: Portuguese-Kabardian pairs (75th percentile similarity: 0.8807)
- rus: Russian-Kabardian pairs (75th percentile similarity: 0.8916)
- spa: Spanish-Kabardian pairs (75th percentile similarity: 0.8718)
- tur: Turkish-Kabardian pairs (75th percentile similarity: 0.8554)

## Dataset Creation

### Source Data

The source data comes from [Tatoeba](https://tatoeba.org/), a free collaborative online database of example sentences.

#### Initial Data Collection and Normalization

1. Sentences were extracted from Tatoeba's database
2. Filtered by language pairs
3. Processed and verified using NLLB-200 model
4. Similarity scores were calculated for each translation pair
5. For quality assurance, translations were filtered using the 75th percentile of similarity scores for each language pair as a threshold

### Annotations

The similarity scores were computed using the NLLB-200 model to ensure translation quality.

## Additional Information

### Dataset Curators

This dataset was curated by [Your Name/Organization]

### Licensing Information

This dataset is licensed under [Creative Commons Attribution 2.0](https://creativecommons.org/licenses/by/2.0/).

### Citation Information

```bibtex
@misc{tatoeba_translations_2024,
    title={Multilingual Tatoeba Translations Dataset},
    author={Your Name},
    year={2024},
    publisher={Hugging Face},
    howpublished={\url{https://huggingface.co/datasets/username/dataset-name}}
}
```

### Contributions

Thanks to [@Tatoeba](https://tatoeba.org/) for providing the source data.

## Usage Examples

```python
from datasets import load_dataset

# Load the entire dataset
dataset = load_dataset("username/dataset-name")

# Load specific language pairs
eng_dataset = load_dataset("username/dataset-name", "eng")
rus_dataset = load_dataset("username/dataset-name", "rus")

# Example of accessing data
for example in eng_dataset['train']:
    source = example['source_text']
    target = example['target_text']
    similarity = example['similarity_score']
    # Your processing code...
```