Update README.md

README.md

@@ -1,4 +1,23 @@
 ---
+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
@@ -57,3 +76,139 @@ configs:
   - 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...
+```