abdullaharoon/Urdu-Multi-Domain-Benchmark

Urdu Multi-Domain Datasets

33 labeled Urdu datasets (288,899 examples) for text classification in Nastaliq (Perso-Arabic) and Roman Urdu (Latin). Each domain is a separate Hub subset so you can download one task at a time.

Authors: Muhammad Abdullah Haroon and Maryam Bashir, FAST-NUCES, Lahore.

Companion paper: Domain Robustness of Multilingual NLP Models Across Urdu and Roman Urdu Scripts.

Permanent archive: Zenodo DOI 10.5281/zenodo.22195610.

How to load

Pick a subset (config) from the list on the Hub, or pass name=:

from datasets import load_dataset

ds = load_dataset("abdullaharoon/Urdu-Multi-Domain-Benchmark", "nastaliq_fake_news_detection_domain_a_multi_topic_news")
print(ds)
print(ds["train"][0])

List every subset:

from datasets import get_dataset_config_names
print(get_dataset_config_names("abdullaharoon/Urdu-Multi-Domain-Benchmark"))

Files can also be downloaded from the Files tab (Nastaliq/ and Roman/ folders) or as a zip from the dataset page.

What's in the collection

Script	Task	Domains	What the labels mean
Nastaliq	Fake News Detection	4	0 = 0, 1 = 1
Nastaliq	Hate Speech	5	0 = normal, 1 = hate
Nastaliq	QA	4	0 = invalid, 1 = valid
Nastaliq	Sentiment Analysis	5	0 = negative, 1 = positive
Roman	Cyber Abuse or Abusive Language	4	0 = normal, 1 = abusive
Roman	Product or eCommerce Reviews	3	Negative / Neutral / Positive
Roman	Sentiment Analysis	4	0 = negative, 1 = positive
Roman	Twitter or Social Media Opinions	4	Negative / Neutral / Positive

Task definitions

• Sentiment analysis: polarity of the input text (negative vs positive).
• Hate speech / cyber abuse: whether the text is abusive or not.
• Fake news detection: whether a news item is labeled real or fake in the source corpus.
• QA pair validation: a question concatenated with a candidate answer; valid if the answer belongs to that question, invalid if it is a swapped distractor. Labels are only invalid / valid (never the full answer string).
• Product reviews / Twitter opinions: three-way polarity (Negative, Neutral, Positive).

Per-dataset sizes

#	Subset (name=)	Script	Task	Domain	Train	Test	Total	Label counts
1	nastaliq_fake_news_detection_domain_a_multi_topic_news	Nastaliq	Fake News Detection	Multi Topic News	638	261	899	0: 400, 1: 499
2	nastaliq_fake_news_detection_domain_b_ax_to_grind	Nastaliq	Fake News Detection	Ax to Grind	8,066	1,746	9,812	0: 4909, 1: 4903
3	nastaliq_fake_news_detection_domain_c_general_pakistani_news	Nastaliq	Fake News Detection	General Pakistani News	18,542	4,629	23,171	0: 12159, 1: 11012
4	nastaliq_fake_news_detection_domain_d_fact_checking_platform	Nastaliq	Fake News Detection	Fact Checking Platform	10,710	2,678	13,388	0: 6711, 1: 6677
5	nastaliq_hate_speech_domain_a_social_media_offensive	Nastaliq	Hate Speech	Social Media Offensive	17,408	4,223	21,631	0: 12785, 1: 8846
6	nastaliq_hate_speech_domain_b_social_media_hate_speech	Nastaliq	Hate Speech	Social Media Hate Speech	20,941	3,986	24,927	0: 12030, 1: 12897
7	nastaliq_hate_speech_domain_c_politics_sports_health_family	Nastaliq	Hate Speech	Politics Sports Health Family	1,360	340	1,700	0: 600, 1: 1100
8	nastaliq_hate_speech_domain_d_sports_labor_arts_education	Nastaliq	Hate Speech	Sports Labor Arts Education	1,360	340	1,700	0: 850, 1: 850
9	nastaliq_hate_speech_domain_e_inter_faith_sectarian_ethnic	Nastaliq	Hate Speech	Inter Faith Sectarian Ethnic	17,407	4,222	21,629	0: 13034, 1: 8595
10	nastaliq_qa_domain_a_religious_hadith	Nastaliq	QA	Religious Hadith	1,312	268	1,580	0: 820, 1: 760
11	nastaliq_qa_domain_b_general_knowledge	Nastaliq	QA	General Knowledge	9,304	2,002	11,306	0: 5805, 1: 5501
12	nastaliq_qa_domain_c_long_form_qna	Nastaliq	QA	Long Form QnA	896	224	1,120	0: 560, 1: 560
13	nastaliq_qa_domain_d_short_form_qna	Nastaliq	QA	Short Form QnA	2,768	692	3,460	0: 1730, 1: 1730
14	nastaliq_sentiment_analysis_domain_a_twitter_social_media	Nastaliq	Sentiment Analysis	Twitter Social Media	783	196	979	0: 499, 1: 480
15	nastaliq_sentiment_analysis_domain_b_movie_reviews	Nastaliq	Sentiment Analysis	Movie Reviews	7,356	1,807	9,163	0: 4537, 1: 4626
16	nastaliq_sentiment_analysis_domain_c_civic_social_topics	Nastaliq	Sentiment Analysis	Civic Social Topics	129	33	162	0: 81, 1: 81
17	nastaliq_sentiment_analysis_domain_d_governance_political	Nastaliq	Sentiment Analysis	Governance Political	685	293	978	0: 499, 1: 479
18	nastaliq_sentiment_analysis_domain_e_socioeconomic_agricultural	Nastaliq	Sentiment Analysis	Socioeconomic Agricultural	2,000	382	2,382	0: 1188, 1: 1194
19	roman_cyber_abuse_or_abusive_language_domain_a_youtube_comments	Roman	Cyber Abuse or Abusive Language	YouTube Comments	4,003	865	4,868	0: 2415, 1: 2453
20	roman_cyber_abuse_or_abusive_language_domain_b_content_creator_comments	Roman	Cyber Abuse or Abusive Language	Content Creator Comments	1,040	260	1,300	0: 580, 1: 720
21	roman_cyber_abuse_or_abusive_language_domain_c_social_media_diverse_vocab	Roman	Cyber Abuse or Abusive Language	Social Media Diverse Vocab	1,040	260	1,300	0: 600, 1: 700
22	roman_cyber_abuse_or_abusive_language_domain_d_general_online_video	Roman	Cyber Abuse or Abusive Language	General Online Video	960	240	1,200	0: 550, 1: 650
23	roman_product_or_ecommerce_reviews_domain_a_daraz_ecommerce	Roman	Product or eCommerce Reviews	Daraz Ecommerce	13,592	3,352	16,944	Negative: 4355, Neutral: 2458, Positive: 10131
24	roman_product_or_ecommerce_reviews_domain_b_restaurant_food	Roman	Product or eCommerce Reviews	Restaurant Food	600	150	750	Negative: 42, Neutral: 117, Positive: 591
25	roman_product_or_ecommerce_reviews_domain_d_electronics_gaming_delivery	Roman	Product or eCommerce Reviews	Electronics Gaming Delivery	22,471	4,195	26,666	Negative: 9742, Neutral: 6325, Positive: 10599
26	roman_sentiment_analysis_domain_a_mixed_public_discourse	Roman	Sentiment Analysis	Mixed Public Discourse	19,295	2,149	21,444	0: 9979, 1: 11465
27	roman_sentiment_analysis_domain_b_social_media_politics_drama	Roman	Sentiment Analysis	Social Media Politics Drama	8,799	2,194	10,993	0: 5297, 1: 5696
28	roman_sentiment_analysis_domain_c_youtube_entertainment	Roman	Sentiment Analysis	YouTube Entertainment	16,266	3,804	20,070	0: 8405, 1: 11665
29	roman_sentiment_analysis_domain_d_utilities_apps_courses	Roman	Sentiment Analysis	Utilities Apps Courses	8,799	2,194	10,993	0: 5297, 1: 5696
30	roman_twitter_or_social_media_opinions_domain_a_twitter_political_social	Roman	Twitter or Social Media Opinions	Twitter Political Social	14,318	3,566	17,884	Negative: 5294, Neutral: 6899, Positive: 5691
31	roman_twitter_or_social_media_opinions_domain_b_consumer_electronics_gaming	Roman	Twitter or Social Media Opinions	Consumer Electronics Gaming	1,160	290	1,450	Negative: 491, Neutral: 477, Positive: 482
32	roman_twitter_or_social_media_opinions_domain_c_public_civic_services	Roman	Twitter or Social Media Opinions	Public Civic Services	1,160	290	1,450	Negative: 490, Neutral: 483, Positive: 477
33	roman_twitter_or_social_media_opinions_domain_d_mixed_social_commentary	Roman	Twitter or Social Media Opinions	Mixed Social Commentary	1,280	320	1,600	Negative: 549, Neutral: 516, Positive: 535

Total: 33 datasets, 236,448 train + 52,451 test = 288,899 examples.

Schema

Column	Type	Description
text	string	Input document (or Question: ...\nAnswer: ... for QA).
label	class	Short class name only. Binary tasks use 0/1; 3-class review/opinion tasks use Negative/Neutral/Positive.

text,label
ye product bohat acha hai,Positive

Directory layout

Nastaliq/<task>/<domain>/train.csv
Nastaliq/<task>/<domain>/test.csv
Roman/<task>/<domain>/train.csv
Roman/<task>/<domain>/test.csv

Only standardized benchmark splits are distributed in this release. Original third-party source files remain local and are not redistributed; see the provenance table for their source and license.

Sources and construction

Base corpora were collected from public research releases (Hugging Face, Kaggle, GitHub, and published Urdu datasets). Where a domain did not exist at useful scale, additional examples were seed-prompted with Grok from organic examples and manually checked by the authors. The table below records origin and the share of Grok-generated text.

#	Script	Task	Domain	Source	Citation	License	Origin	Grok-generated
1	Nastaliq	Fake News Detection	Multi Topic News	Urdu Fake News Shared Task (FIRE)	Amjad et al. (2020)	CC BY 4.0	Organic (Mainstream Portals)	0%
2	Nastaliq	Fake News Detection	Ax to Grind	Ax-to-Grind Partisan Fake News	Malik et al. (2020)	CC BY-NC 4.0	Organic (Partisan Websites)	0%
3	Nastaliq	Fake News Detection	General Pakistani News	General Pakistani Verified/False News	Malik et al. (2020)	CC BY 4.0	Organic (News Outlets)	0%
4	Nastaliq	Fake News Detection	Fact Checking Platform	Urdu Fact-Check Aggregator	Malik et al. (2020)	CC BY-SA 4.0	Organic (Fact-checking Claims)	0%
5	Nastaliq	Hate Speech	Social Media Offensive	Urdu Offensive Social Media Dataset	Rizwan et al. (2020)	CC BY 4.0	Organic (Twitter / YouTube)	0%
6	Nastaliq	Hate Speech	Social Media Hate Speech	Urdu Hate Speech Corpus	Amjad et al. (2021)	CC BY-NC 4.0	Organic (Twitter / Facebook)	0%
7	Nastaliq	Hate Speech	Politics Sports Health Family	Topical Hate Speech Multi-Domain	Rizwan et al. (2020)	CC BY 4.0	Organic seeds + Grok (manual check)	70%
8	Nastaliq	Hate Speech	Sports Labor Arts Education	Institutional Hate & Harassment	Rizwan et al. (2020)	CC BY 4.0	Organic seeds + Grok (manual check)	70%
9	Nastaliq	Hate Speech	Inter Faith Sectarian Ethnic	Sectarian & Ethnic Hate Speech	Rizwan et al. (2020)	Research Only	Organic (Public Forums / Social)	0%
10	Nastaliq	QA	Religious Hadith	Urdu Hadith QA Corpus	Kundi et al. (2014)	Public Domain / Open	Organic (Authenticated Texts)	0%
11	Nastaliq	QA	General Knowledge	Urdu General Knowledge Trivia	Kundi et al. (2014)	CC BY 4.0	Organic (Encyclopedic)	0%
12	Nastaliq	QA	Long Form QnA	Explanatory Long-Form QA	Kundi et al. (2014)	CC BY 4.0	Organic seeds + Grok (manual check)	60%
13	Nastaliq	QA	Short Form QnA	Factual Short-Form QA	Kundi et al. (2014)	CC BY 4.0	Organic seeds + Grok (manual check)	50%
14	Nastaliq	Sentiment Analysis	Twitter Social Media	Urdu Twitter Sentiment Corpus	Mukund & Srihari (2010)	CC BY-NC 4.0	Organic (Twitter/X)	0%
15	Nastaliq	Sentiment Analysis	Movie Reviews	Urdu Movie Reviews Dataset	Asghar et al. (2019)	Research Only	Organic (Editorial / IMDB Urdu)	0%
16	Nastaliq	Sentiment Analysis	Civic Social Topics	Civic Discourse Corpus	Kundi et al. (2014)	CC BY 4.0	Organic + Grok expansion (manual check)	25%
17	Nastaliq	Sentiment Analysis	Governance Political	Urdu Political Opinions Corpus	Asghar et al. (2019)	CC BY-NC-SA 4.0	Organic (News & Forums)	0%
18	Nastaliq	Sentiment Analysis	Socioeconomic Agricultural	Socioeconomic Urdu Sentiments	Asghar et al. (2019)	CC BY 4.0	Organic + Grok expansion (manual check)	20%
19	Roman	Cyber Abuse or Abusive Language	YouTube Comments	Roman Urdu Cyberbullying Dataset	Ghulam et al. (2019)	CC BY-NC 4.0	Organic (YouTube Comments)	0%
20	Roman	Cyber Abuse or Abusive Language	Content Creator Comments	Creator Targeted Callouts	Ghulam et al. (2019)	CC BY 4.0	Organic seeds + Grok (manual check)	65%
21	Roman	Cyber Abuse or Abusive Language	Social Media Diverse Vocab	Diverse Slur & Harassment Lexicon	Ghulam et al. (2019)	CC BY 4.0	Organic seeds + Grok (manual check)	65%
22	Roman	Cyber Abuse or Abusive Language	General Online Video	Video Discussion Harassment	Ghulam et al. (2019)	CC BY 4.0	Organic seeds + Grok (manual check)	65%
23	Roman	Product or eCommerce Reviews	Daraz Ecommerce	Daraz Code-Mixed Product Reviews	Alam et al. (2021)	CC BY 4.0	Organic (Daraz.pk)	0%
24	Roman	Product or eCommerce Reviews	Restaurant Food	Kababjees & Restaurant Reviews	Alam et al. (2021)	CC BY-NC 4.0	Organic (Restaurant Reviews)	0%
25	Roman	Product or eCommerce Reviews	Electronics Gaming Delivery	Roman Urdu Customer Feedback & Delivery	Alam et al. (2021)	CC BY 4.0	Organic (Multi-category Reviews)	0%
26	Roman	Sentiment Analysis	Mixed Public Discourse	Roman Urdu Sentiment (RUS-20K)	Sharf & Rahman (2017)	CC BY 4.0	Organic (Social Media)	0%
27	Roman	Sentiment Analysis	Social Media Politics Drama	Roman Urdu Drama / Politics Comments	Alam et al. (2021)	CC BY-NC 4.0	Organic (Facebook / Twitter)	0%
28	Roman	Sentiment Analysis	YouTube Entertainment	Roman Urdu YouTube Sentiment 27K	Alam et al. (2021)	CC BY 4.0	Organic (YouTube Comments)	0%
29	Roman	Sentiment Analysis	Utilities Apps Courses	Roman Urdu App & Course Reviews	Sharf & Rahman (2017)	CC BY 4.0	Organic (Play Store & Feedback)	0%
30	Roman	Twitter or Social Media Opinions	Twitter Political Social	Roman Urdu Twitter Opinions (New.csv)	Sharf & Rahman (2017)	CC BY 4.0	Organic (Twitter Political Stream)	0%
31	Roman	Twitter or Social Media Opinions	Consumer Electronics Gaming	Consumer Electronics & Gaming Opinions	Alam et al. (2021)	CC BY 4.0	Organic seeds + Grok (manual check)	70%
32	Roman	Twitter or Social Media Opinions	Public Civic Services	Public Infrastructure & Civic Feedback	Alam et al. (2021)	CC BY 4.0	Organic seeds + Grok (manual check)	70%
33	Roman	Twitter or Social Media Opinions	Mixed Social Commentary	Social Trends Commentary	Alam et al. (2021)	CC BY 4.0	Organic seeds + Grok (manual check)	65%

Split integrity

An exact-match train/test audit removed 4,196 overlapping test rows before this release; the post-cleanup exact overlap is zero. A near-duplicate audit remains a planned follow-up.

License

Standardized splits are released under CC BY-NC 4.0 (non-commercial research). Underlying source corpora keep their original licenses in the table above.

Citation

@article{haroon2026domain,
  title={Domain Robustness of Multilingual NLP Models Across Urdu and Roman Urdu Scripts},
  author={Haroon, Muhammad Abdullah and Bashir, Maryam},
  journal={arXiv preprint},
  year={2026}
}

Contact

• Muhammad Abdullah Haroon — l230734@lhr.nu.edu.pk
• Maryam Bashir — maryam.bashir@lhr.nu.edu.pk
• FAST School of Computing, FAST-NUCES, Lahore, Pakistan