A Twitter dataset of 2,280 user accounts, 368,028 tweets, collected by Yang and Menczer (OSoMe/Indiana University), 2023. Each record contains user_id, label (bot/human), dataset source, full tweet objects with text and metadata. First large-scale dataset of AI-powered social bots (Fox8 botnet) with GPT-generated tweet content; includes human baseline accounts for comparison.. Suitable for AI bot detection, LLM-generated content identification, social botnet analysis..
A X (Twitter) with Community Notes annotations dataset of 152,266 multimodal posts (21,845 AI-generated, 104,915 miscaptioned), collected by Chrysidis, Papadopoulos & Papadopoulos via X Community Notes (2026). Each record contains tweet ID, reply/favorite/view/retweet counts, Community Notes text and timestamp, and misinformation type label (miscaptioned / ai_generated / other). Large-scale dataset of 152,266 multimodal X posts with Community Notes — tracking the rise of AI-generated misinformation with 21,845 confirmed AI-generated posts. Includes engagement metrics showing that AI-generated content achieves disproportionate virality despite slower initial reporting.. Suitable for AI-generated content detection, misinformation classification, Community Notes effectiveness, multimodal misinformation analysis, virality and engagement modeling, temporal degradation of detectors.
A Twitter (OffensEval 2020 — 5 languages) dataset of 61,922 tweets across 5 languages with offensive/non-offensive labels, collected by Zampieri et al. for SemEval-2020 Task 12. Each record contains tweet text, binary offensive label (OFF/NOT), language, split. Combined OffensEval 2020 benchmark across Arabic, Danish, English, Greek, and Turkish — 61,922 tweets from the most-participated SemEval task ever (528 teams). Suitable for multilingual offensive language detection, cross-lingual transfer, content moderation, NLP benchmarking.
A Twitter (OffensEval 2020 — Turkish) dataset of 34,792 Turkish tweets (6,757 offensive, 28,035 non-offensive), collected by Zampieri et al. for SemEval-2020 Task 12. Each record contains tweet text, binary offensive label (OFF/NOT), split (train/test). OffensEval 2020 Turkish — part of the most popular SemEval task of all time (528 teams), enabling multilingual offensive language research. Suitable for offensive language detection, multilingual NLP, cross-lingual transfer learning, content moderation benchmarking.
A Twitter (crowdsourced hate speech annotation) dataset of 24,783 tweets labeled as hate speech (1,430), offensive (19,190), or neither (4,163), collected by Davidson et al. via CrowdFlower crowdsourcing (2017). Each record contains tweet text, majority class label (0=hate/1=offensive/2=neither), per-label coder counts, total coder count. Davidson et al. 2017 is one of the most cited hate speech datasets, with 24,783 tweets annotated by multiple crowdworkers — distinguished by separating hate speech from merely offensive language. Suitable for hate speech detection, offensive language classification, content moderation, NLP benchmarking, crowdsourced annotation research.
A YouTube (with cross-platform spread on Twitter and Facebook) dataset of 7,659 removed YouTube videos with multilingual titles and descriptions, 249,346 associated tweets, and 18,897,277 Facebook shares, collected by Knuutila (2021) via YouTube Data API and CrowdTangle, tracking videos shared on public social media that were subsequently removed by YouTube for COVID-19 misinformation. Each record contains video titles and descriptions in multiple languages, view counts and subscriber counts where recoverable, Facebook reactions/comments/shares, Twitter and Facebook post ID lists for social spread tracking, removal and publication timestamps, and Wayback Machine archive URLs. The dataset is notable for its one of the few public datasets documenting the complete lifecycle of platform-removed health misinformation, linking video metadata to cross-platform social spread and enabling research into both content and dissemination patterns of COVID-19 misinformation. Intended for research into health misinformation detection, content moderation research, cross-platform information spread analysis, COVID-19 infodemic studies, multilingual misinformation narrative analysis, and platform policy impact assessment.
A Twitter (OffensEval 2020 — Greek) dataset of 10,287 Greek tweets (2,728 offensive, 7,559 non-offensive), collected by Zampieri et al. for SemEval-2020 Task 12. Each record contains tweet text, binary offensive label (OFF/NOT), split (train/test). OffensEval 2020 Greek — part of the most popular SemEval task of all time (528 teams), enabling multilingual offensive language research. Suitable for offensive language detection, multilingual NLP, cross-lingual transfer learning, content moderation benchmarking.
A Twitter (OffensEval 2020 — Arabic) dataset of 9,666 Arabic tweets (1,919 offensive, 7,747 non-offensive), collected by Zampieri et al. for SemEval-2020 Task 12. Each record contains tweet text, binary offensive label (OFF/NOT), split (train/test). OffensEval 2020 Arabic — part of the most popular SemEval task of all time (528 teams), enabling multilingual offensive language research. Suitable for offensive language detection, multilingual NLP, cross-lingual transfer learning, content moderation benchmarking.
A Twitter dataset of 4,952 tweets (1,238 rumour, 3,714 non-rumour) from the 2013 Boston Marathon bombings, collected by Zubiaga, Liakata, and Procter — Augmented PHEME (Zenodo, 2019). Each record contains tweet text, binary rumour/non-rumour label, timestamp, reaction thread structure linking source tweets to replies. Augmented extension of PHEME covering the Boston Marathon bombings event with both rumour threads and non-rumour threads, providing a balanced benchmark for rumour detection with reaction graph structure. Suitable for rumour detection, misinformation classification, crisis communication analysis, stance detection, conversational thread modelling.
A Twitter (OffensEval 2020 — English) dataset of 3,887 English tweets (1,080 offensive, 2,807 non-offensive), collected by Zampieri et al. for SemEval-2020 Task 12. Each record contains tweet text, binary offensive label (OFF/NOT), split (train/test). OffensEval 2020 English — part of the most popular SemEval task of all time (528 teams), enabling multilingual offensive language research. Suitable for offensive language detection, multilingual NLP, cross-lingual transfer learning, content moderation benchmarking.
A Twitter (OffensEval 2020 — Danish) dataset of 3,290 Danish tweets (425 offensive, 2,865 non-offensive), collected by Zampieri et al. for SemEval-2020 Task 12. Each record contains tweet text, binary offensive label (OFF/NOT), split (train/test). OffensEval 2020 Danish — part of the most popular SemEval task of all time (528 teams), enabling multilingual offensive language research. Suitable for offensive language detection, multilingual NLP, cross-lingual transfer learning, content moderation benchmarking.
A 4chan, Twitter, and Weibo (cross-platform) dataset of 2,455 multimodal posts (1,318 hate, 1,137 non-hate), collected by Mira AI Lab (2024). Each record contains image filename, image OCR text, post text, hate/non-hate binary label, hate category (gender/race/religion/politics/violence/health/etc), and human-annotated reason for classification. Multimodal cross-platform collection hate speech dataset combining image OCR text and post text with fine-grained category labels across 8 hate dimensions — 54% of posts are hate speech with human-annotated reasoning. Suitable for multimodal hate speech detection, cross-platform analysis, hate category classification, image-text hate speech, content moderation research. The package includes the 2,455 meme images (img/) that the records name.
A Cross-platform meme corpus (4chan, Twitter, Weibo) dataset of 1,318 hate-speech labeled meme records from 3 platforms, collected by Mira AI Lab via cross-platform meme collection. Each record contains image filename, OCR-extracted image text, post caption text, hate speech label, category list (politics/race/religion/etc.), and human-readable annotation reason per record. M3-Hate is the hate-speech filtered subset of the M3 meme dataset, with 1,318 hateful memes across 4chan (1062), Twitter (132), and Weibo (124) — multimodal with both image OCR text and post captions. Suitable for hate speech detection, multimodal meme analysis, cross-platform content moderation, image-text classification, harmful content research. The package includes the 1,318 meme images (img/) alongside the JSON records.
A Twitter/X dataset of 526 multimodal posts (132 hate, 394 non-hate), collected by Mira AI Lab (2024). Each record contains image filename, image OCR text, post text, hate/non-hate binary label, hate category (gender/race/religion/politics/violence/health/etc), and human-annotated reason for classification. Multimodal Twitter subset hate speech dataset combining image OCR text and post text with fine-grained category labels across 8 hate dimensions — 25% of posts are hate speech with human-annotated reasoning. Suitable for multimodal hate speech detection, cross-platform analysis, hate category classification, image-text hate speech, content moderation research. The package includes the 526 meme images (img/) that the records name.
Facebook
YouTube
4chan
Weibo