NumanKaanKaratas/turkish-sentences

Turkish Sentences

Turkish Sentences is a clean, duplicate-free Turkish text corpus prepared for NLP and language-model training workflows. The dataset contains Turkish sentences and short lexical entries built around Turkish roots, word forms, homonyms, and morphology-rich vocabulary.

Dataset Summary

• —Language: Turkish (tr)
• —Format: Parquet
• —Split: train
• —Rows: 1,978,236
• —Schema: one column, text
• —Created: 2026-05-31T19:38:26+00:00
• —Duplicate status: deduplicated
• —Text status: normalized and cleaned

Data Fields

No auxiliary length, token-count, or id columns are included. The Parquet file intentionally keeps the training surface minimal.

Corpus Statistics

Quality Notes

The corpus is published as a clean Parquet dataset with a single text column. Records are normalized, empty entries are excluded, and duplicate texts are removed while preserving the first occurrence order.

Usage

from datasets import load_dataset

dataset = load_dataset("NumanKaanKaratas/turkish-sentences")
print(dataset["train"][0]["text"])

Example Records

Dişi aslanın dişi kırıldı.
Yüzünü yıkayan adamın cebinde yüz lirası vardı.
Yazı yazan adam yaz ayında yazarlık yaptı.
Kır saçlı dede kır çiçeklerini topladı.
Dik yokuşta diktiğim fidan kurumuş.

Intended Uses

This dataset is suitable for Turkish language-model pretraining or continued pretraining, tokenizer evaluation, morphology-aware representation learning, lexical coverage checks, sentence embedding experiments, and Turkish text generation research.

The short lexical entries can be useful for root and word-form coverage, while the longer sentence entries provide contextual Turkish usage.

Limitations

The dataset is a broad text corpus, not a human-annotated benchmark. It does not include labels, task-specific annotations, or a separate validation/test split. Users should create their own evaluation split for supervised experiments.

Citation

If you use this dataset in your research, please cite:

@dataset{karatas2026turkishsentences,
  title     = {Turkish Sentences},
  author    = {Numan Kaan Karataş},
  year      = {2026},
  publisher = {HuggingFace},
  url       = {https://huggingface.co/datasets/NumanKaanKaratas/turkish-sentences}
}

License

This dataset is released under the MIT License.

Reproducibility

Detailed build metadata is included in metadata.json. The published Parquet SHA-256 is:

a4ff72cba2b48ec35b01b1fc90642b2c2072ffefad2f121762da80ad8a8d2a0d