SPEAK-PP/openslr-sinhala-synthetic-spell-errors-quarter

Sinhala Dyslexic Spelling Correction Dataset

Dataset Description

This dataset contains Sinhala and code-mixed (Sinhala-English) text pairs for training spelling correction models, specifically designed to address dyslexia-like spelling errors.

Features

• dyslexic_sentence: Input text with dyslexia-like spelling errors (string)
• correct_sentence: Corrected output text (string)

Dataset Statistics

Split	Samples
Train	37,056
Test	9,265
Total	46,321

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("SPEAK-ASR/openslr-sinhala-synthetic-spell-errors-quarter")

# Access train and test splits
train_data = dataset['train']
test_data = dataset['test']

# Example
print(train_data[0])
# {'dyslexic_sentence': 'මහවවැලි ගඟට ගොස් ආපුස එයන ගමනේදී', 'correct_sentence': 'මහවැලි ගඟට ගොස් ආපසු එන ගමනේදී'}

Dataset Creation

This dataset was created to address spelling correction challenges in Sinhala text, including:

Error Types Included

• Character substitution: Similar-looking Sinhala characters swapped
• Character deletion: Missing characters in words
• Character insertion: Extra characters added
• Character transposition: Swapped adjacent characters
• Phonetic errors: Similar sounding Sinhala characters (e.g., ණ/න, ෂ/ස)
• Diacritic errors: Vowel sign (matra) mistakes
• Word order errors: Shuffled word positions in sentences
• Code-mixed errors: Errors in Sinhala-English mixed text

Languages

• Primary: Sinhala (සිංහල)
• Secondary: English (code-mixed content)

Example Pairs

Dyslexic (Input)	Correct (Output)
මහවවැලි ගඟට ගොස් ආපුස එයන ගමනේදී	මහවැලි ගඟට ගොස් ආපසු එන ගමනේදී
ඓවන් ශ්‍රේෂ්ඨ ජාථියක් බිහි කිරීමට	එවන් ශ්‍රේෂ්ඨ ජාතියක් බිහි කිරීමට
යම්කිසි මනුෂ්‍යයෙක් ෂිල් රකිනවානම්	යම්කිසි මනුෂ්‍යයෙක් සිල් රකිනවානම්

Model Training

This dataset is suitable for fine-tuning:

• T5-based models
• mBART models
• Sequence-to-sequence transformers
• Custom encoder-decoder models

Recommended Preprocessing

def preprocess_function(examples):
    inputs = ["correct: " + text for text in examples['dyslexic_sentence']]
    targets = examples['correct_sentence']
    return {'input': inputs, 'target': targets}

Citation

If you use this dataset, please cite:

@dataset{sinhala_dyslexic_spelling_correction,
  title={Sinhala Dyslexic Spelling Correction Dataset},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/SPEAK-ASR/openslr-sinhala-synthetic-spell-errors-quarter}
}

License

This dataset is released under the MIT License.