Back to home
All Datasets 🏷
All datasets from our datasets repository and community bucket.
Also check out the list of supported Metrics 📉.
37 results
Allocine Dataset: A Large-Scale French Movie Reviews Dataset. This is a dataset for binary sentiment classification, made of user reviews scraped from It contains 100k positive and 100k negative reviews divided into 3 balanced splits: train (160k reviews), val (20k) and test (20k).
Chinese word segmentation (CWS) trained from open source corpus faces dramatic performance drop when dealing with domain text, especially for a domain with lots of special terms and diverse writing styles, such as the biomedical domain. However, building domain-specific CWS requires extremely high annotation cost. In this paper, we propose an approach by exploiting domain-invariant knowledge from high resource to low resource domains. Extensive experiments show that our mode achieves consistently higher accuracy than the single-task CWS and other transfer learning baselines, especially when there is a large disparity between source and target domains. This dataset is the accompanied medical Chinese word segmentation (CWS) dataset. The tags are in BIES scheme. For more details see
The Bengali Hate Speech Dataset is a collection of Bengali articles collected from Bengali news articles, news dump of Bengali TV channels, books, blogs, and social media. Emphasis was placed on Facebook pages and newspaper sources because they attract close to 50 million followers and is a common source of opinions and hate speech. The raw text corpus contains 250 million articles and the full dataset is being prepared for release. This is a subset of the full dataset. This dataset was prepared for hate-speech text classification benchmark on Bengali, an under-resourced language.
Corpus of domain names scraped from Common Crawl and manually annotated to add word boundaries (e.g. "commoncrawl" to "common crawl"). Breaking domain names such as "openresearch" into component words "open" and "research" is important for applications such as Text-to-Speech synthesis and web search. Common Crawl is an open repository of web crawl data that can be accessed and analyzed by anyone. Specifically, we scraped the plaintext (WET) extracts for domain names from URLs that contained diverse letter casing (e.g. "OpenBSD"). Although in the previous example, segmentation is trivial using letter casing, this was not always the case (e.g. "NASA"), so we had to manually annotate the data. The dataset is stored as plaintext file where each line is an example of space separated segments of a domain name. The examples are stored in their original letter casing, but harder and more interesting examples can be generated by lowercasing the input first.
DART is a large and open-domain structured DAta Record to Text generation corpus with high-quality sentence annotations with each input being a set of entity-relation triples following a tree-structured ontology. It consists of 82191 examples across different domains with each input being a semantic RDF triple set derived from data records in tables and the tree ontology of table schema, annotated with sentence description that covers all facts in the triple set. DART is released in the following paper where you can find more details and baseline results:
Data and code from our "Inferring Which Medical Treatments Work from Reports of Clinical Trials", NAACL 2019. This work concerns inferring the results reported in clinical trials from text. The dataset consists of biomedical articles describing randomized control trials (RCTs) that compare multiple treatments. Each of these articles will have multiple questions, or 'prompts' associated with them. These prompts will ask about the relationship between an intervention and comparator with respect to an outcome, as reported in the trial. For example, a prompt may ask about the reported effects of aspirin as compared to placebo on the duration of headaches. For the sake of this task, we assume that a particular article will report that the intervention of interest either significantly increased, significantly decreased or had significant effect on the outcome, relative to the comparator. The dataset could be used for automatic data extraction of the results of a given RCT. This would enable readers to discover the effectiveness of different treatments without needing to read the paper.
A dataset to study Fake News in Portuguese, presenting a supposedly false News along with their respective fact check and classification. The data is collected from the ClaimReview, a structured data schema used by fact check agencies to share their results in search engines, enabling data collect in real time. The FACTCK.BR dataset contains 1309 claims with its corresponding label.
The directory data contains a corpus of Finnish technology related news articles with a manually prepared named entity annotation (digitoday.2014.csv). The text material was extracted from the archives of Digitoday, a Finnish online technology news source ( The corpus consists of 953 articles (193,742 word tokens) with six named entity classes (organization, location, person, product, event, and date). The corpus is available for research purposes and can be readily used for development of NER systems for Finnish.
HEAD-QA is a multi-choice HEAlthcare Dataset. The questions come from exams to access a specialized position in the Spanish healthcare system, and are challenging even for highly specialized humans. They are designed by the Ministerio de Sanidad, Consumo y Bienestar Social. The dataset contains questions about the following topics: medicine, nursing, psychology, chemistry, pharmacology and biology.
PUBHEALTH is a comprehensive dataset for explainable automated fact-checking of public health claims. Each instance in the PUBHEALTH dataset has an associated veracity label (true, false, unproven, mixture). Furthermore each instance in the dataset has an explanation text field. The explanation is a justification for which the claim has been assigned a particular veracity label. The dataset was created to explore fact-checking of difficult to verify claims i.e., those which require expertise from outside of the journalistics domain, in this case biomedical and public health expertise. It was also created in response to the lack of fact-checking datasets which provide gold standard natural language explanations for verdicts/labels. NOTE: There are missing labels in the dataset and we have replaced them with -1.
This repository contains a dump of thousands of public domain works in Hebrew, from Project Ben-Yehuda, in plaintext UTF-8 files, with and without diacritics (nikkud). The metadata (pseudocatalogue.csv) file is a list of titles, authors, genres, and file paths, to help you process the dump. All these works are in the public domain, so you are free to make any use of them, and do not need to ask for permission. There are 10078 files, 3181136 lines
HebrewSentiment is a data set consists of 12,804 user comments to posts on the official Facebook page of Israel’s president, Mr. Reuven Rivlin. In October 2015, we used the open software application Netvizz (Rieder, 2013) to scrape all the comments to all of the president’s posts in the period of June – August 2014, the first three months of Rivlin’s presidency.2 While the president’s posts aimed at reconciling tensions and called for tolerance and empathy, the sentiment expressed in the comments to the president’s posts was polarized between citizens who warmly thanked the president, and citizens that fiercely critiqued his policy. Of the 12,804 comments, 370 are neutral; 8,512 are positive, 3,922 negative. Data Annotation: A trained researcher examined each comment and determined its sentiment value, where comments with an overall positive sentiment were assigned the value 1, comments with an overall negative sentiment were assigned the value -1, and comments that are off-topic to the post’s content were assigned the value 0. We validated the coding scheme by asking a second trained researcher to code the same data. There was substantial agreement between raters (N of agreements: 10623, N of disagreements: 2105, Coehn’s Kappa = 0.697, p = 0).
KILT tasks training and evaluation data. - [FEVER]( | Fact Checking | fever - [AIDA CoNLL-YAGO]( | Entity Linking | aidayago2 - [WNED-WIKI]( | Entity Linking | wned - [WNED-CWEB]( | Entity Linking | cweb - [T-REx]( | Slot Filling | trex - [Zero-Shot RE]( | Slot Filling | structured_zeroshot - [Natural Questions]( | Open Domain QA | nq - [HotpotQA]( | Open Domain QA | hotpotqa - [TriviaQA]( | Open Domain QA | triviaqa - [ELI5]( | Open Domain QA | eli5 - [Wizard of Wikipedia]( | Dialogue | wow To finish linking TriviaQA questions to the IDs provided, follow the instructions [here](
This is a Korean paired question dataset containing labels indicating whether two questions in a given pair are semantically identical. This dataset was used to evaluate the performance of [KoGPT2]( on a phrase detection downstream task.
The Large Spanish Corpus is a compilation of 15 unlabelled Spanish corpora spanning Wikipedia to European parliament notes. Each config contains the data corresponding to a different corpus. For example, "all_wiki" only includes examples from Spanish Wikipedia. By default, the config is set to "combined" which loads all the corpora; with this setting you can also specify the number of samples to return per corpus by configuring the "split" argument.
Machine learning models are trained to find patterns in data. NLP models can inadvertently learn socially undesirable patterns when training on gender biased text. In this work, we propose a general framework that decomposes gender bias in text along several pragmatic and semantic dimensions: bias from the gender of the person being spoken about, bias from the gender of the person being spoken to, and bias from the gender of the speaker. Using this fine-grained framework, we automatically annotate eight large scale datasets with gender information. In addition, we collect a novel, crowdsourced evaluation benchmark of utterance-level gender rewrites. Distinguishing between gender bias along multiple dimensions is important, as it enables us to train finer-grained gender bias classifiers. We show our classifiers prove valuable for a variety of important applications, such as controlling for gender bias in generative models, detecting gender bias in arbitrary text, and shed light on offensive language in terms of genderedness.
NewsQA is a challenging machine comprehension dataset of over 100,000 human-generated question-answer pairs. Crowdworkers supply questions and answers based on a set of over 10,000 news articles from CNN, with answers consisting of spans of text from the corresponding articles.
NumerSense is a new numerical commonsense reasoning probing task, with a diagnostic dataset consisting of 3,145 masked-word-prediction probes. We propose to study whether numerical commonsense knowledge can be induced from pre-trained language models like BERT, and to what extent this access to knowledge robust against adversarial examples is. We hope this will be beneficial for tasks such as knowledge base completion and open-domain question answering.
Fused Head constructions are noun phrases in which the head noun is missing and is said to be "fused" with its dependent modifier. This missing information is implicit and is important for sentence understanding.The missing heads are easily filled in by humans, but pose a challenge for computational models. For example, in the sentence: "I bought 5 apples but got only 4.", 4 is a Fused-Head, and the missing head is apples, which appear earlier in the sentence. This is a crowd-sourced dataset of 10k numerical fused head examples (1M tokens).
A well-structured summarization dataset for the Persian language consists of 93,207 records. It is prepared for Abstractive/Extractive tasks (like cnn_dailymail for English). It can also be used in other scopes like Text Generation, Title Generation, and News Category Classification. It is imperative to consider that the newlines were replaced with the `[n]` symbol. Please interpret them into normal newlines (for ex. `t.replace("[n]", "\n")`) and then use them for your purposes.
PubMedQA is a novel biomedical question answering (QA) dataset collected from PubMed abstracts. The task of PubMedQA is to answer research questions with yes/no/maybe (e.g.: Do preoperative statins reduce atrial fibrillation after coronary artery bypass grafting?) using the corresponding abstracts. PubMedQA has 1k expert-annotated, 61.2k unlabeled and 211.3k artificially generated QA instances. Each PubMedQA instance is composed of (1) a question which is either an existing research article title or derived from one, (2) a context which is the corresponding abstract without its conclusion, (3) a long answer, which is the conclusion of the abstract and, presumably, answers the research question, and (4) a yes/no/maybe answer which summarizes the conclusion. PubMedQA is the first QA dataset where reasoning over biomedical research texts, especially their quantitative contents, is required to answer the questions.
dataset consisting of parsed Parsed ASTs that were used to train and evaluate the DeepSyn tool. The Python programs are collected from GitHub repositories by removing duplicate files, removing project forks (copy of another existing repository) ,keeping only programs that parse and have at most 30'000 nodes in the AST and we aim to remove obfuscated files
Question Answering in Context is a dataset for modeling, understanding, and participating in information seeking dialog. Data instances consist of an interactive dialog between two crowd workers: (1) a student who poses a sequence of freeform questions to learn as much as possible about a hidden Wikipedia text, and (2) a teacher who answers the questions by providing short excerpts (spans) from the text. QuAC introduces challenges not found in existing machine comprehension datasets: its questions are often more open-ended, unanswerable, or only meaningful within the dialog context.
The RONEC (Named Entity Corpus for the Romanian language) dataset contains over 26000 entities in ~5000 annotated sentence, belonging to 16 distinct classes. It represents the first initiative in the Romanian language space specifically targeted for named entity recognition
Here are two different adversaries, each of which uses a different procedure to pick the sentence it adds to the paragraph: AddSent: Generates up to five candidate adversarial sentences that don't answer the question, but have a lot of words in common with the question. Picks the one that most confuses the model. AddOneSent: Similar to AddSent, but just picks one of the candidate sentences at random. This adversary is does not query the model in any way.
ThaiSum is a large-scale corpus for Thai text summarization obtained from several online news websites namely Thairath, ThaiPBS, Prachathai, and The Standard. This dataset consists of over 350,000 article and summary pairs written by journalists.
WinoBias, a Winograd-schema dataset for coreference resolution focused on gender bias. The corpus contains Winograd-schema style sentences with entities corresponding to people referred by their occupation (e.g. the nurse, the doctor, the carpenter).
XOR-TyDi QA brings together for the first time information-seeking questions, open-retrieval QA, and multilingual QA to create a multilingual open-retrieval QA dataset that enables cross-lingual answer retrieval. It consists of questions written by information-seeking native speakers in 7 typologically diverse languages and answer annotations that are retrieved from multilingual document collections. There are three sub-tasks: XOR-Retrieve, XOR-EnglishSpan, and XOR-Full.
Dataset built from pairs of YouTube captions where both 'auto-generated' and 'manually-corrected' captions are available for a single specified language. This dataset labels two-way (e.g. ignoring single-sided insertions) same-length token differences in the `diff_type` column. The `default_seq` is composed of tokens from the 'auto-generated' captions. When a difference occurs between the 'auto-generated' vs 'manually-corrected' captions types, the `correction_seq` contains tokens from the 'manually-corrected' captions.