ChamaraVishwajithRajapaksha/sinhala-question-answer-pairs

Sinhala Question Answer Pairs

A curated Sinhala-language instruction fine-tuning dataset built by combining two open-access multilingual and Sinhala-specific sources. It contains question–answer pairs with English characters removed, retaining full Sinhala script and punctuation.

Dataset Summary

Property	Value
Language	Sinhala (si)
Task	Question Answering / Instruction Tuning
License	Apache 2.0
Columns	input, target

Sources

1. CohereLabs/aya_dataset

The Aya Dataset is a multilingual human-annotated instruction fine-tuning dataset covering 65 languages. Only rows with language_code == "sin" (Sinhala) were extracted. The inputs column was mapped to input and targets to target.

2. ihalage/sinhala-finetune-qa-eli5

A Sinhala QA dataset derived from the ELI5 (Explain Like I'm Five) corpus, with questions and answers translated into Sinhala. The sinhala_question column was mapped to input and sinhala_answer to target.

Dataset Structure

Data Fields

Column	Type	Description
input	string	The question or prompt in Sinhala
target	string	The answer or completion in Sinhala

Data Splits

Split	Rows
train	all rows (single split)

Example

{
  "input": "ලොව විශාලතම විශ්වවිද්‍යාලය කුමක්ද?",
  "target": "නිව්යෝර්ක් විශ්වවිද්‍යාලය"
}

Preprocessing

The following preprocessing steps were applied to both source datasets before merging:

• Filtered to Sinhala-only rows using language_code == "sin" (Aya) and the sinhala_question / sinhala_answer columns (ELI5)
• Removed all ASCII English alphabet characters (a–z, A–Z) using regex [a-zA-Z]
• Retained all punctuation marks, Sinhala Unicode characters (U+0D80–U+0DFF), digits, and whitespace
• Dropped rows where both input and target were empty after cleaning
• Stripped leading and trailing whitespace from all fields

Usage

from datasets import load_dataset

ds = load_dataset("ChamaraVishwajithRajapaksha/sinhala-question-answer-pairs")
print(ds["train"][0])

Fine-tuning example

from datasets import load_dataset

ds = load_dataset("ChamaraVishwajithRajapaksha/sinhala-question-answer-pairs", split="train")

# Format as instruction prompt
def format_prompt(example):
    return {
        "prompt": f"### ප්‍රශ්නය:\n{example['input']}\n\n### පිළිතුර:\n{example['target']}"
    }

ds = ds.map(format_prompt)

Intended Uses

• Fine-tuning large language models (LLMs) for Sinhala question answering
• Instruction tuning of multilingual or Sinhala-specific models
• Evaluation of Sinhala NLP pipelines
• Low-resource language research

Limitations

• English characters have been removed, which may affect rows that contained mixed-language content (e.g., proper nouns or technical terms originally in English)
• Coverage is limited to the topics present in the two source datasets
• No deduplication has been applied across the two sources

License

This dataset is released under the Apache 2.0 License, consistent with both source datasets.

Citation

If you use this dataset, please cite the original sources:

@misc{singh2024aya,
  title={Aya Dataset: An Open-Access Collection for Multilingual Instruction Tuning},
  author={Shivalika Singh and Freddie Vargus and Daniel Dsouza and others},
  year={2024},
  eprint={2402.06619},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}

@dataset{ihalage_sinhala_eli5,
  author    = {ihalage},
  title     = {Sinhala Finetune QA ELI5},
  year      = {2023},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/ihalage/sinhala-finetune-qa-eli5}
}

Author

Created by ChamaraVishwajithRajapaksha