Dataset Card for GLUE

Table of Contents

• Dataset Card for GLUE
  • Table of Contents
  • Dataset Description
    • Dataset Summary
    • Supported Tasks and Leaderboards
      • ax
      • cola
      • mnli
      • mnli_matched
      • mnli_mismatched
      • mrpc
      • qnli
      • qqp
      • rte
      • sst2
      • stsb
      • wnli
    • Languages
  • Dataset Structure
    • Data Instances
      • ax
      • cola
      • mnli
      • mnli_matched
      • mnli_mismatched
      • mrpc
      • qnli
      • qqp
      • rte
      • sst2
      • stsb
      • wnli
    • Data Fields
      • ax
      • cola
      • mnli
      • mnli_matched
      • mnli_mismatched
      • mrpc
      • qnli
      • qqp
      • rte
      • sst2
      • stsb
      • wnli
    • Data Splits
      • ax
      • cola
      • mnli
      • mnli_matched
      • mnli_mismatched
      • mrpc
      • qnli
      • qqp
      • rte
      • sst2
      • stsb
      • wnli
  • Dataset Creation
    • Curation Rationale
    • Source Data
      • Initial Data Collection and Normalization
      • Who are the source language producers?
    • Annotations
      • Annotation process
      • Who are the annotators?
    • Personal and Sensitive Information
  • Considerations for Using the Data
    • Social Impact of Dataset
    • Discussion of Biases
    • Other Known Limitations
  • Additional Information
    • Dataset Curators
    • Licensing Information
    • Citation Information
    • Contributions

Dataset Description

• Homepage: https://gluebenchmark.com/
• Repository: https://github.com/nyu-mll/GLUE-baselines
• Paper: https://arxiv.org/abs/1804.07461
• Leaderboard: https://gluebenchmark.com/leaderboard
• Point of Contact: More Information Needed
• Size of downloaded dataset files: 1.00 GB
• Size of the generated dataset: 240.84 MB
• Total amount of disk used: 1.24 GB

Dataset Summary

GLUE, the General Language Understanding Evaluation benchmark (https://gluebenchmark.com/) is a collection of resources for training, evaluating, and analyzing natural language understanding systems.

Supported Tasks and Leaderboards

The leaderboard for the GLUE benchmark can be found at this address. It comprises the following tasks:

ax

A manually-curated evaluation dataset for fine-grained analysis of system performance on a broad range of linguistic phenomena. This dataset evaluates sentence understanding through Natural Language Inference (NLI) problems. Use a model trained on MulitNLI to produce predictions for this dataset.

cola

The Corpus of Linguistic Acceptability consists of English acceptability judgments drawn from books and journal articles on linguistic theory. Each example is a sequence of words annotated with whether it is a grammatical English sentence.

mnli

The Multi-Genre Natural Language Inference Corpus is a crowdsourced collection of sentence pairs with textual entailment annotations. Given a premise sentence and a hypothesis sentence, the task is to predict whether the premise entails the hypothesis (entailment), contradicts the hypothesis (contradiction), or neither (neutral). The premise sentences are gathered from ten different sources, including transcribed speech, fiction, and government reports. The authors of the benchmark use the standard test set, for which they obtained private labels from the RTE authors, and evaluate on both the matched (in-domain) and mismatched (cross-domain) section. They also uses and recommend the SNLI corpus as 550k examples of auxiliary training data.

mnli_matched

The matched validation and test splits from MNLI. See the "mnli" BuilderConfig for additional information.

mnli_mismatched

The mismatched validation and test splits from MNLI. See the "mnli" BuilderConfig for additional information.

mrpc

The Microsoft Research Paraphrase Corpus (Dolan & Brockett, 2005) is a corpus of sentence pairs automatically extracted from online news sources, with human annotations for whether the sentences in the pair are semantically equivalent.

qnli

The Stanford Question Answering Dataset is a question-answering dataset consisting of question-paragraph pairs, where one of the sentences in the paragraph (drawn from Wikipedia) contains the answer to the corresponding question (written by an annotator). The authors of the benchmark convert the task into sentence pair classification by forming a pair between each question and each sentence in the corresponding context, and filtering out pairs with low lexical overlap between the question and the context sentence. The task is to determine whether the context sentence contains the answer to the question. This modified version of the original task removes the requirement that the model select the exact answer, but also removes the simplifying assumptions that the answer is always present in the input and that lexical overlap is a reliable cue.

qqp

The Quora Question Pairs2 dataset is a collection of question pairs from the community question-answering website Quora. The task is to determine whether a pair of questions are semantically equivalent.

rte

The Recognizing Textual Entailment (RTE) datasets come from a series of annual textual entailment challenges. The authors of the benchmark combined the data from RTE1 (Dagan et al., 2006), RTE2 (Bar Haim et al., 2006), RTE3 (Giampiccolo et al., 2007), and RTE5 (Bentivogli et al., 2009). Examples are constructed based on news and Wikipedia text. The authors of the benchmark convert all datasets to a two-class split, where for three-class datasets they collapse neutral and contradiction into not entailment, for consistency.

sst2

The Stanford Sentiment Treebank consists of sentences from movie reviews and human annotations of their sentiment. The task is to predict the sentiment of a given sentence. It uses the two-way (positive/negative) class split, with only sentence-level labels.

stsb

The Semantic Textual Similarity Benchmark (Cer et al., 2017) is a collection of sentence pairs drawn from news headlines, video and image captions, and natural language inference data. Each pair is human-annotated with a similarity score from 1 to 5.

wnli

The Winograd Schema Challenge (Levesque et al., 2011) is a reading comprehension task in which a system must read a sentence with a pronoun and select the referent of that pronoun from a list of choices. The examples are manually constructed to foil simple statistical methods: Each one is contingent on contextual information provided by a single word or phrase in the sentence. To convert the problem into sentence pair classification, the authors of the benchmark construct sentence pairs by replacing the ambiguous pronoun with each possible referent. The task is to predict if the sentence with the pronoun substituted is entailed by the original sentence. They use a small evaluation set consisting of new examples derived from fiction books that was shared privately by the authors of the original corpus. While the included training set is balanced between two classes, the test set is imbalanced between them (65% not entailment). Also, due to a data quirk, the development set is adversarial: hypotheses are sometimes shared between training and development examples, so if a model memorizes the training examples, they will predict the wrong label on corresponding development set example. As with QNLI, each example is evaluated separately, so there is not a systematic correspondence between a model's score on this task and its score on the unconverted original task. The authors of the benchmark call converted dataset WNLI (Winograd NLI).

Languages

The language data in GLUE is in English (BCP-47 en)

Dataset Structure

Data Instances

ax

• Size of downloaded dataset files: 0.22 MB
• Size of the generated dataset: 0.24 MB
• Total amount of disk used: 0.46 MB

An example of 'test' looks as follows.

{
  "premise": "The cat sat on the mat.",
  "hypothesis": "The cat did not sit on the mat.",
  "label": -1,
  "idx: 0
}

cola

• Size of downloaded dataset files: 0.38 MB
• Size of the generated dataset: 0.61 MB
• Total amount of disk used: 0.99 MB

An example of 'train' looks as follows.

{
  "sentence": "Our friends won't buy this analysis, let alone the next one we propose.",
  "label": 1,
  "id": 0
}

mnli

• Size of downloaded dataset files: 312.78 MB
• Size of the generated dataset: 82.47 MB
• Total amount of disk used: 395.26 MB

An example of 'train' looks as follows.

{
  "premise": "Conceptually cream skimming has two basic dimensions - product and geography.",
  "hypothesis": "Product and geography are what make cream skimming work.",
  "label": 1,
  "idx": 0
}

mnli_matched

• Size of downloaded dataset files: 312.78 MB
• Size of the generated dataset: 3.69 MB
• Total amount of disk used: 316.48 MB

An example of 'test' looks as follows.

{
  "premise": "Hierbas, ans seco, ans dulce, and frigola are just a few names worth keeping a look-out for.",
  "hypothesis": "Hierbas is a name worth looking out for.",
  "label": -1,
  "idx": 0
}

mnli_mismatched

• Size of downloaded dataset files: 312.78 MB
• Size of the generated dataset: 3.91 MB
• Total amount of disk used: 316.69 MB

An example of 'test' looks as follows.

{
  "premise": "What have you decided, what are you going to do?",
  "hypothesis": "So what's your decision?",
  "label": -1,
  "idx": 0
}

mrpc

• Size of downloaded dataset files: ??
• Size of the generated dataset: 1.5 MB
• Total amount of disk used: ??

An example of 'train' looks as follows.

{
  "sentence1": "Amrozi accused his brother, whom he called "the witness", of deliberately distorting his evidence.",
  "sentence2": "Referring to him as only "the witness", Amrozi accused his brother of deliberately distorting his evidence.",
  "label": 1,
  "idx": 0
}

qnli

• Size of downloaded dataset files: ??
• Size of the generated dataset: 28 MB
• Total amount of disk used: ??

An example of 'train' looks as follows.

{
  "question": "When did the third Digimon series begin?",
  "sentence": "Unlike the two seasons before it and most of the seasons that followed, Digimon Tamers takes a darker and more realistic approach to its story featuring Digimon who do not reincarnate after their deaths and more complex character development in the original Japanese.",
  "label": 1,
  "idx": 0
}

qqp

• Size of downloaded dataset files: ??
• Size of the generated dataset: 107 MB
• Total amount of disk used: ??

An example of 'train' looks as follows.

{
  "question1": "How is the life of a math student? Could you describe your own experiences?",
  "question2": "Which level of prepration is enough for the exam jlpt5?",
  "label": 0,
  "idx": 0
}

rte

• Size of downloaded dataset files: ??
• Size of the generated dataset: 1.9 MB
• Total amount of disk used: ??

An example of 'train' looks as follows.

{
  "sentence1": "No Weapons of Mass Destruction Found in Iraq Yet.",
  "sentence2": "Weapons of Mass Destruction Found in Iraq.",
  "label": 1,
  "idx": 0
}

sst2

• Size of downloaded dataset files: ??
• Size of the generated dataset: 4.9 MB
• Total amount of disk used: ??

An example of 'train' looks as follows.

{
  "sentence": "hide new secretions from the parental units",
  "label": 0,
  "idx": 0
}

stsb

• Size of downloaded dataset files: ??
• Size of the generated dataset: 1.2 MB
• Total amount of disk used: ??

An example of 'train' looks as follows.

{
  "sentence1": "A plane is taking off.",
  "sentence2": "An air plane is taking off.",
  "label": 5.0,
  "idx": 0
}

wnli

• Size of downloaded dataset files: ??
• Size of the generated dataset: 0.18 MB
• Total amount of disk used: ??

An example of 'train' looks as follows.

{
  "sentence1": "I stuck a pin through a carrot. When I pulled the pin out, it had a hole.",
  "sentence2": "The carrot had a hole.",
  "label": 1,
  "idx": 0
}

Data Fields

The data fields are the same among all splits.

ax

• premise: a string feature.
• hypothesis: a string feature.
• label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).
• idx: a int32 feature.

cola

• sentence: a string feature.
• label: a classification label, with possible values including unacceptable (0), acceptable (1).
• idx: a int32 feature.

mnli

• premise: a string feature.
• hypothesis: a string feature.
• label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).
• idx: a int32 feature.

mnli_matched

• premise: a string feature.
• hypothesis: a string feature.
• label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).
• idx: a int32 feature.

mnli_mismatched

• premise: a string feature.
• hypothesis: a string feature.
• label: a classification label, with possible values including entailment (0), neutral (1), contradiction (2).
• idx: a int32 feature.

mrpc

• sentence1: a string feature.
• sentence2: a string feature.
• label: a classification label, with possible values including not_equivalent (0), equivalent (1).
• idx: a int32 feature.

qnli

• question: a string feature.
• sentence: a string feature.
• label: a classification label, with possible values including entailment (0), not_entailment (1).
• idx: a int32 feature.

qqp

• question1: a string feature.
• question2: a string feature.
• label: a classification label, with possible values including not_duplicate (0), duplicate (1).
• idx: a int32 feature.

rte

• sentence1: a string feature.
• sentence2: a string feature.
• label: a classification label, with possible values including entailment (0), not_entailment (1).
• idx: a int32 feature.

sst2

• sentence: a string feature.
• label: a classification label, with possible values including negative (0), positive (1).
• idx: a int32 feature.

stsb

• sentence1: a string feature.
• sentence2: a string feature.
• label: a float32 regression label, with possible values from 0 to 5.
• idx: a int32 feature.

wnli

• sentence1: a string feature.
• sentence2: a string feature.
• label: a classification label, with possible values including not_entailment (0), entailment (1).
• idx: a int32 feature.

Data Splits

ax

cola

mnli

mnli_matched

mnli_mismatched

mrpc

More Information Needed

qnli

More Information Needed

qqp

More Information Needed

rte

More Information Needed

sst2

More Information Needed

stsb

More Information Needed

wnli

More Information Needed

Dataset Creation

Curation Rationale

More Information Needed

Source Data

Initial Data Collection and Normalization

More Information Needed

Who are the source language producers?

More Information Needed

Annotations

Annotation process

More Information Needed

Who are the annotators?

More Information Needed

Personal and Sensitive Information

More Information Needed

Considerations for Using the Data

Social Impact of Dataset

More Information Needed

Discussion of Biases

More Information Needed

Other Known Limitations

More Information Needed

Additional Information

Dataset Curators

More Information Needed

Licensing Information

The primary GLUE tasks are built on and derived from existing datasets. We refer users to the original licenses accompanying each dataset.

Citation Information

If you use GLUE, please cite all the datasets you use.

In addition, we encourage you to use the following BibTeX citation for GLUE itself:

@inproceedings{wang2019glue,
  title={{GLUE}: A Multi-Task Benchmark and Analysis Platform for Natural Language Understanding},
  author={Wang, Alex and Singh, Amanpreet and Michael, Julian and Hill, Felix and Levy, Omer and Bowman, Samuel R.},
  note={In the Proceedings of ICLR.},
  year={2019}
}

If you evaluate using GLUE, we also highly recommend citing the papers that originally introduced the nine GLUE tasks, both to give the original authors their due credit and because venues will expect papers to describe the data they evaluate on. The following provides BibTeX for all of the GLUE tasks, except QQP, for which we recommend adding a footnote to this page: https://data.quora.com/First-Quora-Dataset-Release-Question-Pairs

@article{warstadt2018neural,
  title={Neural Network Acceptability Judgments},
  author={Warstadt, Alex and Singh, Amanpreet and Bowman, Samuel R.},
  journal={arXiv preprint 1805.12471},
  year={2018}
}
@inproceedings{socher2013recursive,
  title={Recursive deep models for semantic compositionality over a sentiment treebank},
  author={Socher, Richard and Perelygin, Alex and Wu, Jean and Chuang, Jason and Manning, Christopher D and Ng, Andrew and Potts, Christopher},
  booktitle={Proceedings of EMNLP},
  pages={1631--1642},
  year={2013}
}
@inproceedings{dolan2005automatically,
  title={Automatically constructing a corpus of sentential paraphrases},
  author={Dolan, William B and Brockett, Chris},
  booktitle={Proceedings of the International Workshop on Paraphrasing},
  year={2005}
}
@book{agirre2007semantic,
  editor    = {Agirre, Eneko and M`arquez, Llu'{i}s and Wicentowski, Richard},
  title     = {Proceedings of the Fourth International Workshop on Semantic Evaluations (SemEval-2007)},
  month     = {June},
  year      = {2007},
  address   = {Prague, Czech Republic},
  publisher = {Association for Computational Linguistics},
}
@inproceedings{williams2018broad,
  author    = {Williams, Adina and Nangia, Nikita and Bowman, Samuel R.},
  title = {A Broad-Coverage Challenge Corpus for Sentence Understanding through Inference},
  booktitle = {Proceedings of NAACL-HLT},
  year = 2018
}
@inproceedings{rajpurkar2016squad,
  author = {Rajpurkar, Pranav and Zhang, Jian and Lopyrev, Konstantin and Liang, Percy}
  title = {{SQ}u{AD}: 100,000+ Questions for Machine Comprehension of Text},
  booktitle = {Proceedings of EMNLP}
  year = {2016},
  publisher = {Association for Computational Linguistics},
  pages = {2383--2392},
  location = {Austin, Texas},
}
@incollection{dagan2006pascal,
  title={The {PASCAL} recognising textual entailment challenge},
  author={Dagan, Ido and Glickman, Oren and Magnini, Bernardo},
  booktitle={Machine learning challenges. evaluating predictive uncertainty, visual object classification, and recognising tectual entailment},
  pages={177--190},
  year={2006},
  publisher={Springer}
}
@article{bar2006second,
  title={The second {PASCAL} recognising textual entailment challenge},
  author={Bar Haim, Roy and Dagan, Ido and Dolan, Bill and Ferro, Lisa and Giampiccolo, Danilo and Magnini, Bernardo and Szpektor, Idan},
  year={2006}
}
@inproceedings{giampiccolo2007third,
  title={The third {PASCAL} recognizing textual entailment challenge},
  author={Giampiccolo, Danilo and Magnini, Bernardo and Dagan, Ido and Dolan, Bill},
  booktitle={Proceedings of the ACL-PASCAL workshop on textual entailment and paraphrasing},
  pages={1--9},
  year={2007},
  organization={Association for Computational Linguistics},
}
@article{bentivogli2009fifth,
  title={The Fifth {PASCAL} Recognizing Textual Entailment Challenge},
  author={Bentivogli, Luisa and Dagan, Ido and Dang, Hoa Trang and Giampiccolo, Danilo and Magnini, Bernardo},
  booktitle={TAC},
  year={2009}
}
@inproceedings{levesque2011winograd,
  title={The {W}inograd schema challenge},
  author={Levesque, Hector J and Davis, Ernest and Morgenstern, Leora},
  booktitle={{AAAI} Spring Symposium: Logical Formalizations of Commonsense Reasoning},
  volume={46},
  pages={47},
  year={2011}
}

Contributions

Thanks to @patpizio, @jeswan, @thomwolf, @patrickvonplaten, @mariamabarham for adding this dataset.