said-sinhala-ai-dataset/said-hat

Sinhala-HAT (Human and AI-generated Text)

Sinhala-HAT is the document-level binary-classification subset of SAID (Sinhala AI-text Identification Dataset), the first AI-generated-text detection benchmark for Sinhala. SAID has a second, companion subset — Sinhala-BOUND — for sentence-level authorship-boundary detection in mixed human/AI Wikipedia passages; see ORG/said-bound (or wherever that dataset is published).

Sinhala-HAT is a Sinhala dataset of 8,994 human-written and AI-generated texts across three domains — news, social media, and Wikipedia — built for whole-document AI-generated-text detection. It has 4,497 human/AI pairs: each AI document is either written from a title (Mode A, news and Wikipedia only) or a paraphrase of the paired human document (Mode B, all three domains). AI text comes from Gemini 2.5 Pro, DeepSeek-V3 and GPT-4o. Splits are leakage-free by domain × source id.

from datasets import load_dataset

ds = load_dataset("said-sinhala-ai-dataset/said-hat")            # all three domains, pooled
ds = load_dataset("said-sinhala-ai-dataset/said-hat", "news")
ds = load_dataset("said-sinhala-ai-dataset/said-hat", "social_media")
ds = load_dataset("said-sinhala-ai-dataset/said-hat", "wikipedia")

A note on social media punctuation

The social-media text in this dataset has no punctuation, in both classes. In the human posts as collected, punctuation was essentially absent already — only 7 of 1,500 posts (0.5%) contained any punctuation or symbol character. The AI-generated social-media text, by contrast, was fluently punctuated by the generators (97.1% of 1,500 posts). Publishing that pairing as-is would mean punctuation presence alone almost perfectly separates the two classes in this domain — a trivial shortcut with nothing to do with detecting AI-generated text.

To avoid shipping that shortcut, every punctuation and symbol character has been removed from the social-media text, identically for human and AI documents. word_count is recomputed after removal. News and Wikipedia text is unaffected — this only applies to the social_media config/subset.

This is a property of the released text, not of the original generation prompts: the prompts given to the generators did not request punctuation-free output (see prompts/generation_prompts.md). The gap between the two classes predates the AI generation step and was addressed here as a data-cleaning choice, applied identically to both classes as a matter of principle.

Dataset structure

Config	Train	Validation	Test	Total
news	2,048	460	486	2,994
social_media	2,090	454	456	3,000
wikipedia	2,156	436	408	3,000
default (pooled)	6,294	1,350	1,350	8,994

Every split is exactly 50% human / 50% AI. Splits are assigned by domain × source_id, so a human document and its paired AI document(s) always fall in the same split — there is no pair leakage across train/validation/test.

Columns

Column	Type	Description
text_id	string	Unique row id
text	string	The document. Social media text has had all punctuation/symbol characters removed, in both classes — see the note above.
label	int (0/1)	0 = human-written, 1 = AI-generated
domain	string	news, social_media, or wikipedia
generator	string	human, gemini-2.5-pro, deepseek-v3, or gpt-4o
generation_mode	string	human, mode_a (written from a title), or mode_b (paraphrase of the paired human text)
source_id	string	Id of the underlying source article/post
group	string	{domain}_{source_id}; links a human document to its paired AI document(s)
prompt_id	string	Which prompt template generated this row (see prompts/generation_prompts.md); human for human rows
input_tokens	float, nullable	Tokens in the generation prompt; null for human rows
output_tokens	float, nullable	Tokens in the generated output; null for human rows
word_count	int	Whitespace word count of text
split	string	train, validation, or test

Generation modes by domain

Domain	Mode A (from title)	Mode B (paraphrase)
news	yes	yes
social_media	—	yes only
wikipedia	yes	yes

Social media is Mode B only: every AI social-media document is a paraphrase of a real human post, never written from a title alone.

Source data

• News: NSINA, CC BY-SA 4.0 (Hettiarachchi et al., 2024 — see Citation below).
• Social media: Facebook Decade Corpora, CC BY-SA 4.0 (Wijeratne and de Silva, 2020 — see Citation below).
• Wikipedia: Wikipedia Monthly (Sinhala snapshot), CC BY-SA 4.0. Source articles were restricted to those created and last-modified before 2022-11-30 (ChatGPT's public launch), so the human side of the Wikipedia data predates mass LLM-assisted editing.

Source-quality floors applied before generation: news and Wikipedia source articles needed at least 40 and 100 words respectively; social media source posts needed at least 15.

Licensing

This dataset is released under CC BY-SA 4.0 (see LICENSE). If you adapt this dataset, you must release your adaptation under the same licence and attribute the original sources: Sinhala Wikipedia via Wikipedia Monthly (CC BY-SA 4.0), NSINA news text (CC BY-SA 4.0), and Facebook Decade Corpora social-media text (CC BY-SA 4.0). Citations for all three are under Citation below.

AI-generated text was produced by Gemini 2.5 Pro (Google), DeepSeek-V3 (DeepSeek), and GPT-4o (OpenAI); redistribution of their outputs is subject to those providers' terms.

Known limitations

A classifier trained on this dataset should be evaluated with the following in mind:

• Social media punctuation (addressed above): resolved in this release by removing punctuation from both classes.
• Document length: human news and Wikipedia articles run noticeably longer than their AI counterparts (roughly 1.6× by word count).
• Numeric content: AI text written from a title alone (Mode A) contains far fewer numbers than the paired human text, since the generation prompts instructed the model not to invent dates or figures it could not verify.
• Latin-script content: human Wikipedia text contains substantially more Latin-script content (foreign names, citations, scientific terms) than the AI-generated Wikipedia text, a byproduct of the "write entirely in Sinhala" generation constraint.

None of these fully explain classifier performance on this dataset — see the accompanying paper for the full artifact analysis.

Citation

If you use Sinhala-HAT, please cite the SAID paper (TODO below) along with the three source corpora it draws on:

@misc{said2026,
  title  = {SAID: [TODO full title]},
  author = {[TODO author list]},
  year   = {2026},
  note   = {[TODO venue / arXiv id / URL]}
}

@inproceedings{nsina2024,
  author    = {Hettiarachchi, Hansi and Premasiri, Damith and Uyangodage, Lasitha and Ranasinghe, Tharindu},
  title     = {{NSINA: A News Corpus for Sinhala}},
  booktitle = {The 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)},
  year      = {2024},
  month     = {May}
}

@misc{facebookdecadecorpora2020,
  author    = {Wijeratne, Yudhanjaya and de Silva, Nisansa},
  title     = {Sinhala Language Corpora and Stopwords from a Decade of Sri Lankan Facebook},
  year      = {2020},
  publisher = {LIRNEasia}
}

@misc{wikipedia_kamali_2025,
  author    = {Omar Kamali and Omneity Labs},
  title     = {Wikipedia Monthly},
  year      = {2025},
  url       = {https://huggingface.co/datasets/omarkamali/wikipedia-monthly},
  doi       = {10.57967/hf/6575},
  publisher = {Hugging Face}
}

Contact