PlagiaCorpus

PlagiaCorpus is a bilingual (Arabic–English), adversarial and cross-lingual benchmark for forensic plagiarism and text-reuse detection.

It pairs a corpus of 34,885 source passages with 163,758 queries. Each query is one of the following:

• a transformed version of a source passage, through one of five reuse strategies of increasing difficulty: verbatim copy, character-level obfuscation, two independent LLM paraphrasers, and cross-lingual translation with structural rewriting;
• a Safe (non-reused) passage.

Every reused query is linked to its source passage, and the train/test split is grouped by source passage: no passage contributes queries to both splits. The dataset supports source retrieval (find the passage a query was derived from) and reuse detection (decide whether a query reuses a passage of the corpus).

Text records	198,643 (34,885 corpus passages + 163,758 queries)
Languages	Arabic, English
Query split	147,381 train / 16,377 test (90/10, grouped by source passage)
Labels	148,044 reused (label = 1) / 15,714 Safe (label = 0)
Reuse classes	Class 1–5 + Safe
Format	JSON (list of records per file) and Parquet
DOI	10.57967/hf/11369
Code	https://github.com/riotu-lab/PlagiaCorpus
Paper	PlagiaCorpus: A Bilingual, Adversarial, and Cross-Lingual Dataset for Forensic Plagiarism Detection (submitted to Scientific Data)

Why PlagiaCorpus?

Capability	PAN-PC-11	AraPlagDet 2015	PAWS	PAN 2026	PlagiaCorpus
Multilingual support	✓	✗ (Ar only)	✗ (En only)	✗ (En only)	✓
Arabic coverage	✗	✓	✗	✗	✓
Cross-lingual plagiarism	✓	✗	✗	✗	✓
Adversarial / artificial obfuscation	✓	✓	✓	✓	✓
Complex paraphrasing	✓	✓	✓	✓	✓
LLM-generated plagiarism	✗	✗	✗	✓	✓

Files

File	Records	Content
PlagiaCorpus_corpus.json / .parquet	34,885	Source passages: the reference collection to index for retrieval.
PlagiaCorpus_queries_train.json / .parquet	147,381	Training queries (90%).
PlagiaCorpus_queries_test.json / .parquet	16,377	Test queries (10%).

Each source passage, together with its exact and near duplicates, forms a group, and all queries of a group are in the same split. Groups were assigned to the test split within strata (collection × language; Safe queries separately) until 10% of each stratum's queries was reached (seed 42). No source passage and no identical query text occurs in both splits.

Reuse classes

Class	type values	Reuse type	Generation	Label	Train	Test
Class 1	Class 1 (EN), Class 1 (AR)	Direct copy	The source passage, verbatim	1	31,400	3,485
Class 2	Class 2 (EN), Class 2 (AR)	Algorithmic modification	English: leetspeak e→3, a→@, i→1, o→0. Arabic: tatweel (U+0640) inserted after the first character of every word longer than two characters	1	31,400	3,485
Class 3	Class 3 (EN), Class 3 (AR)	Standard paraphrase	Qwen/Qwen2.5-14B-Instruct, same language	1	29,618	3,313
Class 4	Class 4 (EN), Class 4 (AR)	Complex paraphrase	gemma4:31b, same language	1	29,270	3,224
Class 5	Class 5 (EN->AR), Class 5 (AR->EN)	Cross-lingual reuse	gemma4:26b, translation with light structural paraphrasing	1	11,551	1,298
Safe	Safe (EN), Safe (AR)	Non-reused text	Independently sampled passage; 25% carry the Class 2 obfuscation	0	14,142	1,572

• All LLM generations used temperature 0.7. Top-p was 0.9 for Classes 4 and 5 and for the book part of Class 3. Class 3 was limited to 256 new tokens.
• The models were instructed to output only the rewritten (or translated) text. Residual introductions, notes and markdown were removed during curation.
• The Class 5 prompts explicitly discourage literal translation.
• All prompts are published verbatim in the code repository.

Example

These are the transformations of source passage WIKI_006573, with a Safe query for comparison:

Class	Text
Class 1	a technical or scientific diploma or degree from a university or technikon, involving at least a three-year course of study;six months' practical training in the office of a registered practicing patent attorney;sitting the Patent Board Examination.
Class 2	@ t3chn1c@l 0r sc13nt1f1c d1pl0m@ 0r d3gr33 fr0m @ un1v3rs1ty 0r t3chn1k0n, 1nv0lv1ng @t l3@st @ thr33-y3@r c0urs3 0f study;s1x m0nths' pr@ct1c@l tr@1n1ng 1n th3 0ff1c3 0f @ r3g1st3r3d pr@ct1c1ng p@t3nt @tt0rn3y;s1tt1ng th3 P@t3nt B0@rd Ex@m1n@t10n.
Class 3	To qualify, one must hold a diploma or degree in a technical or scientific field from an accredited institution, which requires at least three years of study. Additionally, six months of hands-on training under a licensed and practicing patent attorney is necessary. Finally, passing the Patent Board Examination is required.
Class 4	A minimum three-year technical or scientific degree from a university or technical college.• Six months of practical experience working under a certified patent attorney.• Completion of the Patent Board Examination.
Class 5 (EN->AR)	-الحصول على دبلوم أو درجة علمية في تخصص تقني أو علمي من جامعة أو معهد تقني، بشرط ألا تقل مدة الدراسة عن ثلاث سنوات؛- إتمام ستة أشهر من التدريب العملي في مكتب وكيل براءات اختراع ممارس ومسجل؛ - اجتياز امتحان مجلس براءات الاختراع.
Safe	When the elements of the codomain of a function are vectors, the function is said to be a vector-valued function. These functions are particularly useful in applications, for example modeling physical properties. …

Data fields

Corpus records (PlagiaCorpus_corpus)

Field	Type	Description
_id	string	Unique passage identifier: WIKI_000000 … for Wikipedia, PD_000000 … for books.
text	string	The source passage.
language	string	en or ar.
domain	string	Source collection: Wikipedia (Wikimedia Foundation), accessed April 2026, The Arabic E-Book Corpus (Hindawi Foundation) or PleIAs English-PD.
source	string	Attribution: the Wikipedia article URL, or a citation of the book collection including the book title.
source_license	string	CC BY-SA 4.0, CC BY 4.0 or Public domain.

Query records (PlagiaCorpus_queries_{train,test})

Field	Type	Description
_id	string	Unique query identifier (Q_000000 …).
text	string	The query text.
type	string	Reuse class and language (see Reuse classes). For Class 5, the translation direction.
reuse_class	string	Class 1 … Class 5, or Safe.
language	string	Language of the query text (en or ar).
source_language	string	Language of the source passage; null for Safe queries.
transformation	string	verbatim, algorithmic, llm_paraphrase, llm_cross_lingual or none. Safe queries with Class 2 obfuscation are marked algorithmic.
label	int	1 = reused (Classes 1–5), 0 = Safe.
source	string	The transformation applied and the attribution of the source text.
domain	string	Source collection (same values as in the corpus).
source_license	string	Licence of the underlying text.
generator_model	string	None (Direct Human Copy) (Class 1), None (Algorithmic Modification) (Class 2 and obfuscated Safe), Qwen/Qwen2.5-14B-Instruct (Class 3), gemma4:31b (Class 4), gemma4:26b (Class 5), or Human (Original Text) (Safe).
original_text_id	string	_id of the source passage in the corpus; null for Safe queries.
quality_flags	list	Automatic flags for LLM outputs: possibly_truncated, much_shorter_than_source.

Use original_text_id, not source, to link a query to its passage.

Example query record:

{
    "_id": "Q_000123",
    "text": "To qualify, one must hold a diploma or degree in a technical or scientific field ...",
    "type": "Class 3 (EN)",
    "reuse_class": "Class 3",
    "language": "en",
    "source_language": "en",
    "transformation": "llm_paraphrase",
    "label": 1,
    "source": "Paraphrased by Qwen/Qwen2.5-14B-Instruct from https://en.wikipedia.org/wiki/...",
    "domain": "Wikipedia (Wikimedia Foundation), accessed April 2026",
    "source_license": "CC BY-SA 4.0",
    "generator_model": "Qwen/Qwen2.5-14B-Instruct",
    "original_text_id": "WIKI_006573",
    "quality_flags": []
}

Dataset statistics

Source passages by repository

Repository	License	Passages	%
English Wikipedia	CC BY-SA 4.0	12,102	34.69
Arabic Wikipedia	CC BY-SA 4.0	6,483	18.58
The Arabic E-Book Corpus (Hindawi Foundation books)	CC BY 4.0	9,999	28.66
PleIAs/English-PD	Public domain	6,301	18.06
Total		34,885	100.00

The corpus contains about 5.27 million words.

Queries per class

Word counts are whitespace-separated tokens.

Class	Count	% of queries	Generator	Mean words	Median	Std
Class 1 (EN)	18,403	11.2	None (direct copy)	167.7	120	151.2
Class 1 (AR)	16,482	10.1	None (direct copy)	132.7	88	137.3
Class 2 (EN)	18,403	11.2	None (rule-based)	167.7	120	151.2
Class 2 (AR)	16,482	10.1	None (rule-based)	132.7	88	137.3
Class 3 (EN)	18,252	11.1	Qwen2.5-14B-Instruct	112.8	106	58.9
Class 3 (AR)	14,679	9.0	Qwen2.5-14B-Instruct	89.3	96	32.2
Class 4 (EN)	17,426	10.6	gemma4:31b	122.3	92	94.2
Class 4 (AR)	15,068	9.2	gemma4:31b	104.1	79	77.7
Class 5 (EN->AR)	5,997	3.7	gemma4:26b	162.4	121	137.9
Class 5 (AR->EN)	6,852	4.2	gemma4:26b	195.6	142	163.8
Safe (EN)	5,530	3.4	None (human text)	104.9	70	101.9
Safe (AR)	10,184	6.2	None (human text)	75.9	56	59.6
Total	163,758	100.0

Coverage of the source passages differs by class:

• Classes 1 and 2 cover every source passage.
• Classes 3 and 4 cover the passages whose generation succeeded and passed curation (32,931 and 32,494 passages).
• Class 5 covers 12,849 passages, sampled by language for each translation direction.

Languages

Language	Records	%
English	103,269	51.99
Arabic	95,374	48.01

Dataset creation

1. Source passages.
   • Wikipedia: English and Arabic articles on a broad range of topics (literature, history, culture, science, law, technology, philosophy, geography and more), retrieved through the MediaWiki API in April 2026 and split into paragraphs. Only paragraphs longer than 200 characters were kept, and section headings were removed.
   • Books: paragraphs longer than 150 characters, streamed from PleIAs/English-PD and The Arabic E-Book Corpus.
2. Classes 1–3: generated for every source passage, within its own component.
3. Safe queries (negatives): drawn from text pools kept separate from the source passages: Wikipedia control pools of articles on topics not used for the sources, and later portions of the book collections.
4. Merging: the two components were merged into one corpus with stable _ids.
5. Classes 4 and 5: Class 4 paraphrases were generated for the merged corpus. Class 5 translations were generated for source passages sampled separately by language.
6. Curation:
   • Language filtering: passages not in English or Arabic were removed with their queries (Lingua and an English function-word check).
   • LLM output cleaning: introductions, notes and markdown were stripped; refusals and outputs in the wrong language or script were removed.
   • De-duplication: exact duplicates, and LaBSE cross-lingual near-duplicates at cosine ≥ 0.85, were removed or grouped.
   • Safe obfuscation: 25% of Safe queries per language received the Class 2 rules.
7. Split: grouped 90/10 split, as described above.

The full pipeline, from crawling to curation and statistics, is in the code repository.

Quality: human annotation

A balanced sample of 300 queries (50 per class) was rated independently by two annotators. The annotators did not see the true class.

The annotators judged three things:

• whether the query reuses the passage shown (Is_Plagiarism);
• its Linguistic Quality and Semantic Preservation;
• its Obfuscation Difficulty.

The last three were rated on ordinal scales. Agreement was measured with Cohen's κ, quadratic-weighted for the ordinal dimensions.

Dimension	N	Cohen's κ	Weighted κ	Within ±1
Is_Plagiarism	300	0.967	–	–
Linguistic Quality	300	0.485	0.730	88.67%
Semantic Preservation	299	0.627	0.921	94.31%
Obfuscation Difficulty	258	0.411	0.739	95.74%

N excludes items marked as not applicable. The sample was drawn before curation. The annotated sheets and the agreement script are in the code repository.

How to use

Load with 🤗 Datasets

from datasets import load_dataset

repo = "riotu-lab/PlagiaCorpus"

queries = load_dataset(repo, "queries")                  # DatasetDict: train, test
corpus = load_dataset(repo, "corpus", split="corpus")    # 34,885 passages

passages = dict(zip(corpus["_id"], corpus["text"]))
q = queries["test"][0]
print(q["type"], q["label"], q["text"][:200])
if q["original_text_id"] is not None:
    print("source:", passages[q["original_text_id"]][:200])

Suggested tasks

Task	Input	Target	Metrics
Source retrieval	Query, plus the corpus as the index	original_text_id (for label = 1)	Recall@K, MRR@10
Reuse detection	Query with retrieved candidate passages	label	Precision, recall, F1, MCC
Cross-lingual retrieval	Class 5 queries	original_text_id	Recall@K, MRR@10
Robustness analysis	Any of the above	broken down by type, transformation, domain or generator_model	per-class scores

• Index PlagiaCorpus_corpus with your retriever (BM25, dense embeddings, FAISS, ChromaDB, and so on).
• Fine-tune on the train queries and report results on the test queries.
• Safe queries have no source in the corpus. They measure false positives: a good system should not flag them as reused.
• For stricter evaluation, queries with non-empty quality_flags can be excluded.

Limitations

• Synthetic reuse: Classes 2–5 are produced by rules or LLMs, and may not cover every strategy a human plagiarist uses. LLM outputs can occasionally contain errors or omissions.
• Length: Class 3 generation was capped at 256 new tokens, so Class 3 queries are on average shorter than their sources; possibly truncated outputs are flagged in quality_flags.
• Uneven class coverage: Classes 3–5 do not cover every source passage (see the per-class table).
• Noise in public-domain books: some book passages contain OCR errors.

Content disclaimer

PlagiaCorpus was collected automatically from Wikipedia and openly licensed book collections. It may therefore contain political, ethical, cultural, religious or other viewpoints expressed in the source material. Such content is included only as part of that material and does not represent the views of the authors, their institutions or the funding organisations.

License

PlagiaCorpus is released under CC BY-SA 4.0, the license required by its Wikipedia content. The other sources are compatible:

Source	License
Wikipedia text	CC BY-SA 4.0
The Arabic E-Book Corpus (Hallberg 2025)	CC BY 4.0
PleIAs/English-PD	Public domain

Every record carries its attribution in source and its licence in source_license. When reusing the data, attribute PlagiaCorpus and the original sources, and release adapted versions under the same licence.

Citation

If you use PlagiaCorpus, please cite this paper:

@article{ammar2026plagiacorpus,
  title         = {PlagiaCorpus: A Bilingual, Adversarial, and Cross-Lingual Dataset for Forensic Plagiarism Detection},
  author        = {Ammar, Adel and Sibaee, Serry and Boulila, Wadii},
  journal       = {arXiv preprint arXiv:2611.02940},
  year          = {2026},
  doi           = {10.48550/arXiv.2611.02940},
  url           = {https://arxiv.org/abs/2611.02940}
}

Acknowledgments

This research was conducted with the support of the Cultural Research Grants Program launched by the Ministry of Culture in the Kingdom of Saudi Arabia. The authors acknowledge the support of Prince Sultan University.