said-sinhala-ai-dataset/said-bound

Sinhala-BOUND (Sinhala Boundary Detection)

Sinhala-BOUND is the sentence-level boundary-detection subset of SAID (Sinhala AI-text Identification Dataset), the first AI-generated-text detection benchmark for Sinhala. Its companion subset, Sinhala-HAT, covers whole-document detection.

Sinhala-BOUND has 4,244 Sinhala Wikipedia passages in which some sentences are human-written and some were rewritten by an LLM. Every sentence is labelled (39,358 labels, 36.4% AI). The task is to find where authorship changes. AI sentences replace human sentences in place: no sentence is added or removed. Because of this, every mixed passage has an exactly aligned all-human original, which is released alongside it.

from datasets import load_dataset

ds = load_dataset("said-sinhala-ai-dataset/said-bound")                     # all mixed passages
ds = load_dataset("said-sinhala-ai-dataset/said-bound", "continuation")     # one construction type
ds = load_dataset("said-sinhala-ai-dataset/said-bound", "human_originals")  # aligned all-human twins
ds = load_dataset("said-sinhala-ai-dataset/said-bound", "unrelated_human")  # human-only passages

Configs

Config	What it contains	Train	Validation	Test	Total
default	All mixed passages	2,490	836	918	4,244
continuation	Type 1: human prefix, then an AI continuation (one boundary)	965	342	368	1,675
internal_span	Type 2: one internal span of 1–3 sentences rewritten by the model (two boundaries)	796	272	299	1,367
multiple_spans	Type 3: two or three separate AI spans (up to six boundaries)	729	222	251	1,202
human_originals	The all-human source window of every mixed passage	1,517	535	578	2,630
unrelated_human	All-human windows from articles that no mixed passage uses	4,027	669	618	5,314

Splits are assigned by source article (70/15/15) before generation. All passages from one article, and its human original, are in the same split. unrelated_human windows come from articles that no mixed passage uses, so they share no text with any mixed passage.

Generators

Generator	Role	Train	Validation	Test	Total
gpt-4o	seen	1,304	283	311	1,898
deepseek-v3	seen	1,186	273	302	1,761
gemini-2.5-pro	held out	0	280	305	585

Gemini 2.5 Pro writes validation and test passages only, so it can be held out entirely as an unseen generator. To measure transfer to an unseen generator, train on the two seen generators, tune on their validation passages, and report Gemini test passages separately (generator_role).

Why the human originals matter

Every passage in the mixed configs contains at least one AI sentence. A tagger trained and scored only on such passages never sees a document without a boundary. In the accompanying paper, an XLM-R tagger trained this way flags AI content in 95% of fully human passages, at every decision threshold. Scores on mixed passages cannot reveal this failure.

The two human-only configs make it measurable:

• human_originals is the exact counterfactual "twin" of each mixed passage: the same sentences, in the same order, before any were replaced. Any AI flag on a twin is an unambiguous error. Join on window_id to pair a twin with its mixed passages (passage_ids lists them).
• unrelated_human is a false-alarm test set with no content link to any mixed passage. It also supplies human-only training documents that are not content-matched to any mixed passage.

We recommend reporting the false-alarm rate on fully human passages alongside the scores on mixed passages.

Columns: mixed configs

Column	Type	Description
passage_id	string	Unique passage id
text	string	The passage as one string
sentences	list[string]	The passage split into sentences
labels	list[int]	One label per sentence: 0 = human-written, 1 = AI-generated
boundaries	list[int]	Sentence indices where authorship changes; i means the change is between sentence i-1 and sentence i
ai_spans	list[list[int]]	AI spans as half-open [start, end) sentence ranges
construction_type	string	continuation, internal_span or multiple_spans
generator	string	gpt-4o, deepseek-v3 or gemini-2.5-pro
generator_role	string	seen (in training) or held_out (validation and test only)
window_id	string	Id of the source window; links the passage to its human original
source_id	string	Id of the source Wikipedia article
title	string	Article title
url	string	Article URL on Sinhala Wikipedia
n_sentences	int	Number of sentences (6–12)
n_words	int	Whitespace word count
ai_sentence_ratio	float	Share of AI sentences in the passage
prompt_id	string	Prompt template used (see prompts/generation_prompts.md)
split	string	train, validation or test

Columns: human_originals and unrelated_human

The columns are the same as in the mixed configs where they apply: window_id, source_id, title, url, text, sentences, labels (all 0), n_sentences, n_words and split. human_originals also has passage_ids, the mixed passages built from that window.

How it was built

• Source text. A Sinhala Wikipedia snapshot taken in 2022, before ChatGPT's public release. Articles were cleaned from the raw wikitext with an eleven-step filter (namespaces, redirects, disambiguation pages, stubs, list pages, prose length and script checks), leaving 9,390 articles. These were cut into windows of 6 to 12 consecutive sentences.

• Generation. A model rewrites selected sentences in the context of the passage, and the rewritten sentences replace the originals. The human text around every boundary is therefore unchanged. Type 1 switch points are sampled between 20% and 80% of the passage.

• Generators. GPT-4o (openai/gpt-4o-2024-11-20), DeepSeek-V3 (deepseek/deepseek-chat) and Gemini 2.5 Pro (google/gemini-2.5-pro), all called through OpenRouter with temperature 0.8 and top-p 0.95. Mistral Nemo was piloted and rejected: only 9.2% of its outputs passed validation, against 55–67% for the other three.

• Validation. Every generated span passes seven automatic checks:

  • exact sentence count;
  • word count within 25% of the replaced text;
  • no preamble;
  • no Markdown;
  • Sinhala script;
  • not a copy of the original;
  • numbers unchanged.

  Failing outputs were regenerated, never hand-edited.

• Alignment. For every passage, each human sentence was checked to equal the sentence at the same position in its human original.

The prompts are in prompts/generation_prompts.md.

Known limitations

• Wikipedia only. Boundary detection is not covered for news or social media.
• One held-out generator. Transfer results on Gemini 2.5 Pro show that transfer is imperfect. They do not predict performance on other model families.
• Surface formatting. Human Wikipedia sentences sometimes carry formatting that generator output lacks: a space before punctuation, parentheses, or a missing final punctuation mark. In the accompanying paper, normalising these cues in both classes at test time changed exact-boundary F1 by at most 1.2 points.
• Boundary-only training. As described above, a model trained only on the mixed configs learns that every document contains a boundary. Use the human-only configs to measure and reduce this.

Licensing

This dataset is released under CC BY-SA 4.0 (see LICENSE). The human text comes from Sinhala Wikipedia (CC BY-SA 4.0). If you adapt this dataset, you must release your adaptation under the same licence and attribute Sinhala Wikipedia; the url column links every passage to its source article.

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

Citation

@misc{said2026,
  title  = {SAID: A Sinhala AI-Generated Text Dataset and Counterfactual Twin Training for Authorship Boundary Localisation},
  author = {[TODO author list]},
  year   = {2026},
  note   = {[TODO venue / arXiv id / URL]}
}

Contact