CHECKED, a Chinese COVID-19 fake news dataset from the microblog platform Weibo (Yang, Zhou and Zafarani, Social Network Analysis and Mining 2021): 2,104 microblogs posted between 2019-12-31 and 2020-08-30, 344 labelled fake (each with the expert analysis that identified it as false) and 1,760 real, with the 1,185,701 comments and 1,868,174 reposts (forwards) they received, including their text, date and hashed author. Microblog IDs and user IDs are replaced by 32-character hashes and user names are not released.
The puebla mexico earthquake 2017 event of HumAID: 2,015 tweets, each with its ID, its text and one of 8 humanitarian categories assigned by crowdsourcing workers (984 rescue volunteering or donation effort, 367 sympathy and support, 254 injured or dead people, 167 infrastructure and utility damage, 109 other relevant information, 61 requests or urgent needs, 38 not humanitarian, 35 caution and advice), split by upstream into 1,410 train, 205 dev and 400 test tweets.
The Hebrew part of the TextDetox multilingual toxicity dataset, compiled for the PAN 2024 and PAN 2025 TextDetox shared tasks: 2,011 short texts, each labelled toxic or non-toxic (807 toxic, 1,204 non-toxic). For Hebrew the dataset card names the Hebrew Offensive Language Dataset as the source. The file has two columns, the text and the binary label, and no identifiers, dates or author information. 1,977 of the texts are distinct; the rest occur more than once.
The pakistan earthquake 2019 event of HumAID: 1,958 tweets, each with its ID, its text and one of 8 humanitarian categories assigned by crowdsourcing workers (823 sympathy and support, 401 injured or dead people, 213 not humanitarian, 154 other relevant information, 152 rescue volunteering or donation effort, 125 infrastructure and utility damage, 71 caution and advice, 19 requests or urgent needs), split by upstream into 1,370 train, 199 dev and 389 test tweets.
The Dutch data of Task 1 (check-worthiness) of the CLEF 2024 CheckThat! Lab: 1,913 Dutch texts, each labelled by whether it contains a claim worth fact-checking. The labels are Yes (823) and No (1,090). The data come in three files: train 995 (405 Yes, 590 No), dev 252 (102 Yes, 150 No), dev-test 666 (316 Yes, 350 No). Each row has a tweet ID, the tweet's URL, the tweet text and the label.
The midwestern us floods 2019 event of HumAID: 1,880 tweets, each with its ID, its text and one of 7 humanitarian categories assigned by crowdsourcing workers (788 rescue volunteering or donation effort, 389 not humanitarian, 273 other relevant information, 165 sympathy and support, 140 infrastructure and utility damage, 79 caution and advice, 46 requests or urgent needs), split by upstream into 1,316 train, 191 dev and 373 test tweets.
1,800 Chinese-language comments from Bilibili videos about population aging, pensions and low fertility. Each comment has its like count, the video it came from, in-vivo glosses in Chinese and English, the codes from two coders, an agreed final code and a process-model stage. The package also includes metadata for the source videos and the codebook. Zhangwei Zheng (Sunway University) deposited it on Harvard Dataverse under CC0 as replication data for a study of Bilibili comment discourse on aging.
The Portuguese training data of Task 2 (claim normalization) of the CLEF 2025 CheckThat! Lab: 1,735 social media posts, each paired with a normalized claim, a short simplified statement of the claim the post makes. Each row has the post and its normalized claim. The task asks systems to generate the normalized claim from the noisy post. The upstream repository also holds development and test files for this task; this package holds the training file only.
The hurricane matthew 2016 event of HumAID: 1,654 tweets, each with its ID, its text and one of 9 humanitarian categories assigned by crowdsourcing workers (395 sympathy and support, 328 other relevant information, 326 rescue volunteering or donation effort, 224 injured or dead people, 178 infrastructure and utility damage, 76 not humanitarian, 53 requests or urgent needs, 38 displaced people and evacuations, 36 caution and advice), split by upstream into 1,157 train, 168 dev and 329 test tweets.
The Italian training data of the Task 2 (subjectivity in news articles) of the CLEF 2023 CheckThat! Lab: 1,613 sentences from news articles, each labelled subjective (SUBJ) or objective (OBJ). The label counts are OBJ (1,231), SUBJ (382). Each row has a sentence ID, the sentence and the label. The upstream data folder also holds other files of the task (development and test files); this package holds the training file only.
The Italian training data of the Task 2 (subjectivity in news articles) of the CLEF 2024 CheckThat! Lab: 1,613 sentences from news articles, each labelled subjective (SUBJ) or objective (OBJ). The label counts are OBJ (1,231), SUBJ (382). Each row has a sentence ID, the sentence and the label. The upstream data folder also holds other files of the task (development, dev-test and test files); this package holds the training file only.
The Italian training data of the Task 1 (subjectivity in news articles) of the CLEF 2025 CheckThat! Lab: 1,613 sentences from news articles, each labelled subjective (SUBJ) or objective (OBJ). The label counts are OBJ (1,231), SUBJ (382). Each row has a sentence ID, the sentence and the label. The upstream data folder also holds other files of the task (development and dev-test files); this package holds the training file only.
The ecuador earthquake 2016 event of HumAID: 1,563 tweets, each with its ID, its text and one of 8 humanitarian categories assigned by crowdsourcing workers (555 injured or dead people, 394 rescue volunteering or donation effort, 319 sympathy and support, 91 requests or urgent needs, 81 other relevant information, 70 infrastructure and utility damage, 30 caution and advice, 23 not humanitarian), split by upstream into 1,094 train, 159 dev and 310 test tweets.
The content of genai.stackexchange.com (https://genai.stackexchange.com/), a Stack Exchange question-and-answer site, as published in the Stack Exchange Data Dump of April 2024 on the Internet Archive. The posts table holds 1,317 rows: 642 posts (268 questions, 320 answers, 27 tag wiki excerpts and 27 tag wikis) and 675 comments, written between 17 July 2023 and 31 March 2024. Beside it are 3,437 user accounts, 2,733 votes, 2,882 badges awarded, 80 tags, 2,005 post revisions and 48 links between posts. Every answer keeps the question it answers and every comment the post it was written on, so the threads can be rebuilt. The Internet Archive item describes the dump as anonymised; the votes table names the voter in only 5 rows.
The greece wildfires 2018 event of HumAID: 1,515 tweets, each with its ID, its text and one of 9 humanitarian categories assigned by crowdsourcing workers (495 injured or dead people, 356 rescue volunteering or donation effort, 322 sympathy and support, 159 other relevant information, 74 not humanitarian, 38 infrastructure and utility damage, 26 caution and advice, 25 requests or urgent needs, 20 missing or found people), split by upstream into 1,060 train, 154 dev and 301 test tweets.
The Russian meteor event of CrisisLexT26 (Chelyabinsk, Russia): 1,442 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 8,392 tweet IDs that the CrisisLex authors collected for the event, posted between 14 February 2013 and 7 March 2013, with a flag for the 8,365 that fall within the period they analyzed. For informativeness, the workers judged 642 related and informative, 491 related but not informative, 271 not related and 38 not applicable of the labeled tweets. The record of the event gives its category (Natural, Others, Meteorite), its location and the keywords used to collect the tweets.
The Costa Rica earthquake event of CrisisLexT26 (Costa Rica, Costa Rica): 1,412 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 2,369 tweet IDs that the CrisisLex authors collected for the event, posted between 4 September 2012 and 21 September 2012, with a flag for the 2,193 that fall within the period they analyzed. For informativeness, the workers judged 742 related and informative, 167 related but not informative, 399 not related and 104 not applicable of the labeled tweets. The record of the event gives its category (Natural, Geophysical, Earthquake), its location and the keywords used to collect the tweets.
The 4chan part of M3: 1,400 memes of M3 (image, OCR text, post text, label and rationale), 1,062 labeled hate and 338 normal. Hate categories: race (319), politics (251), violence (178), international relations (152), gender (144), health status (101), religion (88), public health (78).
NoisyHate: 1,339 comments, each as an original sentence (clean_version) paired with a human-written perturbed version meant to evade toxicity detection, with a quality score for the perturbation, the comment's time, reaction counts and toxicity and identity scores, created by Yiran Ye, Thai Le and Dongwon Lee for the ICWSM 2025 paper "NoisyHate: Mining and Curating Human-Written Perturbations for Benchmarking Content Moderation Models". No comment IDs, user names or links are included.
The hateful memes of M3: 1,318 memes of M3 (image, OCR text, post text, label and rationale), 1,318 labeled hate and 0 normal. Hate categories: politics (348), race (334), violence (297), international relations (180), gender (172), religion (108), health status (106), public health (85).
The Savar building collapse event of CrisisLexT26 (Savar, Bangladesh): 1,250 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 4,126 tweet IDs that the CrisisLex authors collected for the event, posted between 23 April 2013 and 1 June 2013, with a flag for the 4,070 that fall within the period they analyzed. For informativeness, the workers judged 836 related and informative, 75 related but not informative, 305 not related and 34 not applicable of the labeled tweets. The record of the event gives its category (Human-induced, Unintentional, Collapse), its location and the keywords used to collect the tweets.
The italy earthquake aug 2016 event of HumAID: 1,201 tweets, each with its ID, its text and one of 6 humanitarian categories assigned by crowdsourcing workers (579 sympathy and support, 312 rescue volunteering or donation effort, 174 injured or dead people, 54 infrastructure and utility damage, 52 other relevant information, 30 requests or urgent needs), split by upstream into 840 train, 122 dev and 239 test tweets.
The Colorado wildfires event of CrisisLexT26 (Colorado, US): 1,200 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 4,182 tweet IDs that the CrisisLex authors collected for the event, posted between 8 June 2012 and 9 July 2012, with a flag for the 4,172 that fall within the period they analyzed. For informativeness, the workers judged 685 related and informative, 268 related but not informative, 238 not related and 9 not applicable of the labeled tweets. The record of the event gives its category (Natural, Climatological, Wildfire), its location and the keywords used to collect the tweets.
The Queensland Floods event of CrisisLexT26 (Queensland, Australia): 1,200 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 1,223 tweet IDs that the CrisisLex authors collected for the event, posted between 17 January 2013 and 5 February 2013, with a flag for the 1,219 that fall within the period they analyzed. For informativeness, the workers judged 728 related and informative, 191 related but not informative, 261 not related and 20 not applicable of the labeled tweets. The record of the event gives its category (Natural, Hydrological, Floods), its location and the keywords used to collect the tweets.
The Australia wildfires event of CrisisLexT26 (New South Wales, Australia): 1,199 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 1,997 tweet IDs that the CrisisLex authors collected for the event, posted between 12 October 2013 and 3 November 2013, with a flag for the 1,982 that fall within the period they analyzed. For informativeness, the workers judged 704 related and informative, 245 related but not informative, 242 not related and 8 not applicable of the labeled tweets. The record of the event gives its category (Natural, Climatological, Wildfire), its location and the keywords used to collect the tweets.
The Arabic training data of the Task 2 (subjectivity in news articles) of the CLEF 2023 CheckThat! Lab: 1,185 sentences from news articles, each labelled subjective (SUBJ) or objective (OBJ). The label counts are OBJ (905), SUBJ (280). Each row has a sentence ID, the sentence and the label. The upstream data folder also holds other files of the task (development and test files); this package holds the training file only.
The Arabic training data of the Task 2 (subjectivity in news articles) of the CLEF 2024 CheckThat! Lab: 1,185 sentences from news articles, each labelled subjective (SUBJ) or objective (OBJ). The label counts are OBJ (905), SUBJ (280). Each row has a sentence ID, the sentence and the label. The upstream data folder also holds other files of the task (development, dev-test and test files); this package holds the training file only.
The French training data of Task 2 (claim normalization) of the CLEF 2025 CheckThat! Lab: 1,174 social media posts, each paired with a normalized claim, a short simplified statement of the claim the post makes. Each row has the post and its normalized claim. The task asks systems to generate the normalized claim from the noisy post. The upstream repository also holds development and test files for this task; this package holds the training file only.
Every submission and comment that the Voat dataset of Mekacher and Papasavva holds for the Voat subverse v/PizzagateNoMod: 321 submissions and 553 comments written between 16 June 2017 and 9 July 2020, with the author's user name, upvote and downvote counts, the submission each comment belongs to and each comment's depth in its thread. Voat (voat.co) was a news aggregator in the style of Reddit that shut down on 25 December 2020; its communities were called subverses. The package also holds 42 user profiles (registration date and the subverses the user moderated or created) of the users who wrote in this subverse or moderate or own it, and the profile of the subverse (description, subscriber count, creation date).
The Glasgow helicopter crash event of CrisisLexT26 (Glasgow, UK): 1,100 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 2,563 tweet IDs that the CrisisLex authors collected for the event, posted between 29 November 2013 and 31 December 2013, with a flag for the 2,551 that fall within the period they analyzed. For informativeness, the workers judged 578 related and informative, 340 related but not informative, 177 not related and 5 not applicable of the labeled tweets. The record of the event gives its category (Human-induced, Unintentional, Crash), its location and the keywords used to collect the tweets.
The Hindi training data of Task 2 (claim normalization) of the CLEF 2025 CheckThat! Lab: 1,081 social media posts, each paired with a normalized claim, a short simplified statement of the claim the post makes. Each row has the post and its normalized claim. The task asks systems to generate the normalized claim from the noisy post. The upstream repository also holds development and test files for this task; this package holds the training file only.
The Guatemala earthquake event of CrisisLexT26 (Guatemala, Guatemala): 1,050 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 3,285 tweet IDs that the CrisisLex authors collected for the event, posted between 6 November 2012 and 27 November 2012, with a flag for the 3,261 that fall within the period they analyzed. For informativeness, the workers judged 754 related and informative, 186 related but not informative, 108 not related and 2 not applicable of the labeled tweets. The record of the event gives its category (Natural, Geophysical, Earthquake), its location and the keywords used to collect the tweets.
The Typhoon Yolanda event of CrisisLexT26 (Phillipines, Phillipines): 1,048 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 39,033 tweet IDs that the CrisisLex authors collected for the event, posted between 3 November 2013 and 1 January 2014, with a flag for the 38,881 that fall within the period they analyzed. For informativeness, the workers judged 765 related and informative, 175 related but not informative, 71 not related and 37 not applicable of the labeled tweets. The record of the event gives its category (Natural, Meteorological, Typhoon), its location and the keywords used to collect the tweets.
The LA Airport Shootings event of CrisisLexT26 (Los Angeles, US): 1,032 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 2,746 tweet IDs that the CrisisLex authors collected for the event, posted between 1 November 2013 and 14 November 2013, with a flag for the 2,730 that fall within the period they analyzed. For informativeness, the workers judged 682 related and informative, 230 related but not informative, 87 not related and 33 not applicable of the labeled tweets. The record of the event gives its category (Human-induced, Intentional, Shootings), its location and the keywords used to collect the tweets.
The Italy earthquakes event of CrisisLexT26 (Modena, Italy): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 7,404 tweet IDs that the CrisisLex authors collected for the event, posted between 18 May 2012 and 21 June 2012, with a flag for the 7,351 that fall within the period they analyzed. For informativeness, the workers judged 627 related and informative, 313 related but not informative, 50 not related and 10 not applicable of the labeled tweets. The record of the event gives its category (Natural, Geophysical, Earthquake), its location and the keywords used to collect the tweets.
The Philipinnes Floods event of CrisisLexT26 (Philipinnes, Philipinnes): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 2,950 tweet IDs that the CrisisLex authors collected for the event, posted between 1 August 2012 and 15 August 2012, with a flag for the 2,917 that fall within the period they analyzed. For informativeness, the workers judged 761 related and informative, 145 related but not informative, 88 not related and 6 not applicable of the labeled tweets. The record of the event gives its category (Natural, Hydrological, Floods), its location and the keywords used to collect the tweets.
The Typhoon Pablo event of CrisisLexT26 (Phillipines, Phillipines): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 1,977 tweet IDs that the CrisisLex authors collected for the event, posted between 24 November 2012 and 16 December 2012, with a flag for the 1,944 that fall within the period they analyzed. For informativeness, the workers judged 678 related and informative, 229 related but not informative, 68 not related and 25 not applicable of the labeled tweets. The record of the event gives its category (Natural, Meteorological, Typhoon), its location and the keywords used to collect the tweets.
The Venezuela refinery explosion event of CrisisLexT26 (Falcon, Venezuela): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 2,845 tweet IDs that the CrisisLex authors collected for the event, posted between 24 August 2012 and 11 September 2012, with a flag for the 2,736 that fall within the period they analyzed. For informativeness, the workers judged 546 related and informative, 393 related but not informative, 58 not related and 3 not applicable of the labeled tweets. The record of the event gives its category (Human-induced, Unintentional, Explosion), its location and the keywords used to collect the tweets.
The Alberta Floods event of CrisisLexT26 (Alberta, Canada): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 5,887 tweet IDs that the CrisisLex authors collected for the event, posted between 20 June 2013 and 16 July 2013, with a flag for the 5,789 that fall within the period they analyzed. For informativeness, the workers judged 685 related and informative, 298 related but not informative, 14 not related and 3 not applicable of the labeled tweets. The record of the event gives its category (Natural, Hydrological, Floods), its location and the keywords used to collect the tweets.
The Bohol earthquake event of CrisisLexT26 (Phillipines, Phillipines): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 2,214 tweet IDs that the CrisisLex authors collected for the event, posted between 14 October 2013 and 25 October 2013, all of which fall within the period they analyzed. For informativeness, the workers judged 425 related and informative, 544 related but not informative and 31 not related of the labeled tweets. The record of the event gives its category (Natural, Geophysical, Earthquake), its location and the keywords used to collect the tweets.
The Boston Bombings event of CrisisLexT26 (Boston, US): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 157,636 tweet IDs that the CrisisLex authors collected for the event, posted between 15 April 2013 and 19 June 2013, with a flag for the 157,454 that fall within the period they analyzed. For informativeness, the workers judged 417 related and informative, 512 related but not informative, 19 not related and 52 not applicable of the labeled tweets. The record of the event gives its category (Human-induced, Intentional, Bombings), its location and the keywords used to collect the tweets.
The Brazil nightclub fire event of CrisisLexT26 (Santa Maria, Brazil): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 4,787 tweet IDs that the CrisisLex authors collected for the event, posted between 27 January 2013 and 12 February 2013, all of which fall within the period they analyzed. For informativeness, the workers judged 442 related and informative, 510 related but not informative, 43 not related and 5 not applicable of the labeled tweets. The record of the event gives its category (Human-induced, Unintentional, Fire), its location and the keywords used to collect the tweets.
The Colorado Floods event of CrisisLexT26 (Colorado, US): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 1,778 tweet IDs that the CrisisLex authors collected for the event, posted between 8 September 2013 and 1 October 2013, with a flag for the 1,752 that fall within the period they analyzed. For informativeness, the workers judged 768 related and informative, 157 related but not informative, 70 not related and 5 not applicable of the labeled tweets. The record of the event gives its category (Natural, Hydrological, Floods), its location and the keywords used to collect the tweets.
The Lac-Megantic train crash event of CrisisLexT26 (Lac-Megantic, Quebec, Canada): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 2,497 tweet IDs that the CrisisLex authors collected for the event, posted between 6 July 2013 and 26 July 2013, with a flag for the 2,342 that fall within the period they analyzed. For informativeness, the workers judged 559 related and informative, 407 related but not informative, 31 not related and 3 not applicable of the labeled tweets. The record of the event gives its category (Human-induced, Unintentional, Derailment), its location and the keywords used to collect the tweets.
The Manila Floods event of CrisisLexT26 (Manila, Phillipines): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 2,033 tweet IDs that the CrisisLex authors collected for the event, posted between 17 August 2013 and 28 August 2013, with a flag for the 2,032 that fall within the period they analyzed. For informativeness, the workers judged 628 related and informative, 293 related but not informative, 47 not related and 32 not applicable of the labeled tweets. The record of the event gives its category (Natural, Hydrological, Floods), its location and the keywords used to collect the tweets.
The NYC train crash event of CrisisLexT26 (New York City, US): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 1,084 tweet IDs that the CrisisLex authors collected for the event, posted between 1 December 2013 and 12 December 2013, with a flag for the 1,066 that fall within the period they analyzed. For informativeness, the workers judged 904 related and informative, 95 related but not informative and 1 not applicable of the labeled tweets. The record of the event gives its category (Human-induced, Unintentional, Derailment), its location and the keywords used to collect the tweets.
The Sardinia Floods event of CrisisLexT26 (Sardinia, Italy): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 1,144 tweet IDs that the CrisisLex authors collected for the event, posted between 16 November 2013 and 29 November 2013, with a flag for the 1,143 that fall within the period they analyzed. For informativeness, the workers judged 631 related and informative, 295 related but not informative, 68 not related and 6 not applicable of the labeled tweets. The record of the event gives its category (Natural, Hydrological, Floods), its location and the keywords used to collect the tweets.
The Singapore Haze event of CrisisLexT26 (Singapore, Singapore): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 3,664 tweet IDs that the CrisisLex authors collected for the event, posted between 14 June 2013 and 11 July 2013, with a flag for the 3,639 that fall within the period they analyzed. For informativeness, the workers judged 461 related and informative, 472 related but not informative, 46 not related and 21 not applicable of the labeled tweets. The record of the event gives its category (Mixed, Others, Haze), its location and the keywords used to collect the tweets.
The Spain train crash event of CrisisLexT26 (Santiago de compostela, Spain): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 3,681 tweet IDs that the CrisisLex authors collected for the event, posted between 24 July 2013 and 7 August 2013, all of which fall within the period they analyzed. For informativeness, the workers judged 729 related and informative, 262 related but not informative, 6 not related and 3 not applicable of the labeled tweets. The record of the event gives its category (Human-induced, Unintentional, Derailment), its location and the keywords used to collect the tweets.
The West Texas Explosion event of CrisisLexT26 (West/Waco Texas, US): 1,000 tweets about it, each labeled by crowdsourcing workers for informativeness, information type and information source, and the list of 14,599 tweet IDs that the CrisisLex authors collected for the event, posted between 17 April 2013 and 18 May 2013, with a flag for the 14,505 that fall within the period they analyzed. For informativeness, the workers judged 472 related and informative, 439 related but not informative, 65 not related and 24 not applicable of the labeled tweets. The record of the event gives its category (Human-induced, Unintentional, Explosion), its location and the keywords used to collect the tweets.
Weibo
Twitter (X)
Stack Exchange
4chan
Voat