dimuthulk/sinhala-text-detoxification

📊 Dataset Card for Sinhala Text Detoxification Dataset

📝 Dataset Description

• Repository: dimuthulk/sinhala-text-detoxification
• Language(s) (NLP): Sinhala (si)
• License: Apache 2.0

📋 Dataset Summary

This dataset serves as the primary generative dataset for the second phase of the "Sinhala Offensive Text Detoxification Pipeline" research project conducted at the University of Kelaniya (Electronics and computer science degree program).

It contains parallel text data designed to train Sequence-to-Sequence (Seq2Seq) models to transform offensive or toxic Sinhala language into polite, neutral alternatives without altering the core semantic meaning.

🚀 Uses

✅ Direct Use

This dataset is strictly built for training, validating, and testing Natural Language Generation (NLG) models for text detoxification in the Sinhala language.

❌ Out-of-Scope Use

This dataset must not be used to train generative models to produce toxic, offensive, or discriminatory content.

🗂️ Dataset Structure

📂 Data Splits

The overall dataset consists of 31,870 parallel sentences[cite: 1], professionally split into three distinct sets for optimal model training[cite: 1]:

• Train Set: 25,496 rows (80%) for training the model[cite: 1].
• Validation Set: 3,187 rows (10%) for tuning and performance evaluation during training[cite: 1].
• Test Set: 3,187 rows (10%) for final accuracy and evaluation[cite: 1].

🏷️ Data Fields

(Note: Please update these field names if your CSV column names are different)

• offensive_text (string): The source text containing offensive or toxic Sinhala language.
• detoxified_text (string): The target text containing the neutralized, non-offensive version.

🛠️ Dataset Creation

🔍 Source Data & Curation

The data was collected, merged, and manually verified by the research team to ensure high-quality parallel sentence pairs. The dataset combines individual contributions into a single, robust structure for text-to-text generation tasks.

ℹ️ Additional Information

🎓 Dataset Curators

Research Team - Electronics and computer science, University of Kelaniya (Including Dimuthu Rathnayaka and team members).

📚 Citation Information

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

@misc{sinhala_text_detoxification_2026,
  author = {Rathnayaka, Dimuthu and Team},
  title = {Sinhala Text Detoxification Dataset},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face repository},
  howpublished = {\url{[https://huggingface.co/datasets/dimuthulk/sinhala-text-detoxification](https://huggingface.co/datasets/dimuthulk/sinhala-text-detoxification)}}
}