annotations_creators:
- other
language_creators:
- other
multilinguality:
- monolingual
source_datasets:
- original
paperswithcode_id: superglue
arxiv: 1905.00537
pretty_name: SuperGLUE Benchmark Datasets
tags:
- superglue
- nlp
- benchmark
license: mit
language:
- en
dataset_info:
- config_name: boolq
features:
- name: question
dtype: string
- name: passage
dtype: string
- name: idx
dtype: int64
- name: label
dtype: bool
splits:
- name: train
num_bytes: 6136774
num_examples: 9427
- name: validation
num_bytes: 2103781
num_examples: 3270
- name: test
num_bytes: 2093385
num_examples: 3245
download_size: 6439045
dataset_size: 10333940
- config_name: cb
features:
- name: premise
dtype: string
- name: hypothesis
dtype: string
- name: label
dtype: string
- name: idx
dtype: int64
splits:
- name: train
num_bytes: 89859
num_examples: 250
- name: validation
num_bytes: 22480
num_examples: 56
- name: test
num_bytes: 93492
num_examples: 250
download_size: 137099
dataset_size: 205831
- config_name: copa
features:
- name: premise
dtype: string
- name: choice1
dtype: string
- name: choice2
dtype: string
- name: question
dtype: string
- name: label
dtype: int64
- name: idx
dtype: int64
splits:
- name: train
num_bytes: 50833
num_examples: 400
- name: validation
num_bytes: 12879
num_examples: 100
- name: test
num_bytes: 61846
num_examples: 500
download_size: 84158
dataset_size: 125558
- config_name: multirc
features:
- name: idx
dtype: int64
- name: version
dtype: float64
- name: passage
struct:
- name: questions
list:
- name: answers
list:
- name: idx
dtype: int64
- name: label
dtype: int64
- name: text
dtype: string
- name: idx
dtype: int64
- name: question
dtype: string
- name: text
dtype: string
splits:
- name: train
num_bytes: 2393721
num_examples: 456
- name: validation
num_bytes: 429255
num_examples: 83
- name: test
num_bytes: 858870
num_examples: 166
download_size: 2053244
dataset_size: 3681846
- config_name: record
features:
- name: source
dtype: string
- name: passage
struct:
- name: entities
list:
- name: end
dtype: int64
- name: start
dtype: int64
- name: text
dtype: string
- name: qas
list:
- name: answers
list:
- name: end
dtype: int64
- name: start
dtype: int64
- name: text
dtype: string
- name: idx
dtype: int64
- name: query
dtype: string
- name: idx
dtype: int64
splits:
- name: train
num_bytes: 110591940
num_examples: 65709
- name: validation
num_bytes: 12375907
num_examples: 7481
- name: test
num_bytes: 11509574
num_examples: 7484
download_size: 71256085
dataset_size: 134477421
- config_name: rte
features:
- name: premise
dtype: string
- name: hypothesis
dtype: string
- name: label
dtype: string
- name: idx
dtype: int64
splits:
- name: train
num_bytes: 877041
num_examples: 2490
- name: validation
num_bytes: 94010
num_examples: 277
- name: test
num_bytes: 973916
num_examples: 3000
download_size: 1269005
dataset_size: 1944967
- config_name: wic
features:
- name: word
dtype: string
- name: sentence1
dtype: string
- name: sentence2
dtype: string
- name: idx
dtype: int64
- name: label
dtype: bool
- name: start1
dtype: int64
- name: start2
dtype: int64
- name: end1
dtype: int64
- name: end2
dtype: int64
- name: version
dtype: float64
splits:
- name: train
num_bytes: 767620
num_examples: 5428
- name: validation
num_bytes: 94651
num_examples: 638
- name: test
num_bytes: 207006
num_examples: 1400
download_size: 591526
dataset_size: 1069277
- config_name: wsc
features:
- name: text
dtype: string
- name: target
struct:
- name: span1_index
dtype: int64
- name: span1_text
dtype: string
- name: span2_index
dtype: int64
- name: span2_text
dtype: string
- name: idx
dtype: int64
- name: label
dtype: bool
splits:
- name: train
num_bytes: 91597
num_examples: 554
- name: validation
num_bytes: 21950
num_examples: 104
- name: test
num_bytes: 32011
num_examples: 146
download_size: 47100
dataset_size: 145558
configs:
- config_name: boolq
data_files:
- split: train
path: boolq/train-*
- split: validation
path: boolq/validation-*
- split: test
path: boolq/test-*
- config_name: cb
data_files:
- split: train
path: cb/train-*
- split: validation
path: cb/validation-*
- split: test
path: cb/test-*
- config_name: copa
data_files:
- split: train
path: copa/train-*
- split: validation
path: copa/validation-*
- split: test
path: copa/test-*
- config_name: multirc
data_files:
- split: train
path: multirc/train-*
- split: validation
path: multirc/validation-*
- split: test
path: multirc/test-*
- config_name: record
data_files:
- split: train
path: record/train-*
- split: validation
path: record/validation-*
- split: test
path: record/test-*
- config_name: rte
data_files:
- split: train
path: rte/train-*
- split: validation
path: rte/validation-*
- split: test
path: rte/test-*
- config_name: wic
data_files:
- split: train
path: wic/train-*
- split: validation
path: wic/validation-*
- split: test
path: wic/test-*
- config_name: wsc
data_files:
- split: train
path: wsc/train-*
- split: validation
path: wsc/validation-*
- split: test
path: wsc/test-*
SuperGLUE Benchmark Datasets
This repository contains the SuperGLUE benchmark datasets. Each dataset is available as a separate configuration, making it easy to load individual datasets using the datasets library.
Dataset Descriptions
Datasets Included
BoolQ: A question-answering task where each example consists of a short passage and a yes/no question about the passage. The questions are provided anonymously and unsolicited by users of the Google search engine and paired with a paragraph from a Wikipedia article containing the answer.
CB (CommitmentBank): A natural language inference task where each example consists of a premise containing an embedded clause, and the corresponding hypothesis is the extraction of that clause. The task focuses on determining whether the premise entails the hypothesis, contradicts it, or is neutral.
COPA (Choice of Plausible Alternatives): A causal reasoning task where the system selects the more plausible alternative between two choices given a premise. Each example consists of a premise and two possible alternatives, and the task is to choose the alternative that has a causal relationship with the premise.
MultiRC (Multiple Sentence Reading Comprehension): A reading comprehension task where each example consists of a context paragraph, a question about the paragraph, and a list of possible answers. The task requires identifying all correct answers for each question, and there may be multiple correct answers.
ReCoRD (Reading Comprehension with Commonsense Reasoning Dataset): A cloze-style reading comprehension task that evaluates a model’s ability to use commonsense reasoning to predict which entity is missing from a passage. Each example consists of a passage with a missing entity and a list of possible entities to fill in the blank.
RTE (Recognizing Textual Entailment): A textual entailment task that involves determining whether a given premise entails a hypothesis. Each example consists of a premise and a hypothesis, and the task is to predict whether the hypothesis is true based on the premise.
WiC (Word-in-Context): A word sense disambiguation task that determines if a word is used in the same sense in two different contexts. Each example consists of two sentences containing the same word, and the task is to decide whether the word has the same meaning in both sentences.
WSC (Winograd Schema Challenge): A pronoun resolution task where the system must determine the antecedent of a pronoun in a sentence. Each example consists of a sentence with a pronoun and a list of possible antecedents, and the task is to select the correct antecedent.
Each dataset has been preprocessed to ensure consistency across train, validation, and test splits. Missing keys in the test split have been filled with dummy values (type-aware) to match the features found in the training and validation splits.
Languages
All data in SuperGLUE is in English.
Dataset Structure
Data Instances
Each task in SuperGLUE is split into train, validation, and test sets. An example instance from each configuration is provided in the task-specific sections below.
BoolQ
- train split: 9427 examples
- validation split: 3270 examples
- test split: 3245 examples
An example of 'train' looks as follows.
{
"question": "do iran and afghanistan speak the same language",
"passage": "Persian language -- Persian (/\u02c8p\u025c\u02d0r\u0292\u0259n, -\u0283\u0259n/), also known by its endonym Farsi (\u0641\u0627\u0631\u0633\u06cc f\u0101rsi (f\u0252\u02d0\u027e\u02c8si\u02d0) ( listen)), is one of the Western Iranian languages within the Indo-Iranian branch of the Indo-European language family. It is primarily spoken in Iran, Afghanistan (officially known as Dari since 1958), and Tajikistan (officially known as Tajiki since the Soviet era), and some other regions which historically were Persianate societies and considered part of Greater Iran. It is written in the Persian alphabet, a modified variant of the Arabic script, which itself evolved from the Aramaic alphabet.",
"idx": 0,
"label": true
}
CB
- train split: 250 examples
- validation split: 56 examples
- test split: 250 examples
An example of 'train' looks as follows.
{
"premise": "It was a complex language. Not written down but handed down. One might say it was peeled down.",
"hypothesis": "the language was peeled down",
"label": "entailment",
"idx": 0
}
COPA
- train split: 400 examples
- validation split: 100 examples
- test split: 500 examples
An example of 'train' looks as follows.
{
"premise": "My body cast a shadow over the grass.",
"choice1": "The sun was rising.",
"choice2": "The grass was cut.",
"question": "cause",
"label": 0,
"idx": 0
}
MultiRC
- train split: 456 examples
- validation split: 83 examples
- test split: 166 examples
An example of 'train' looks as follows.
{
"idx": 0,
"version": 1.1,
"passage": {
"questions": [
{
"answers": [
{
"idx": 0,
"label": 0,
"text": "Children, Gerd, or Dorian Popa"
},
{
"idx": 1,
"label": 0,
"text": "Monetary rewards"
},
{
"idx": 2,
"label": 1,
"text": "Asking Pakistan to help the USA"
},
{
"idx": 3,
"label": 1,
"text": "Meeting with General Musharraf"
},
{
"idx": 4,
"label": 1,
"text": "President Clinton offered the moon in terms of better relations with the United States"
},
{
"idx": 5,
"label": 0,
"text": "A Presidential visit in March"
},
{
"idx": 6,
"label": 1,
"text": "Paper checks"
},
{
"idx": 7,
"label": 0,
"text": "Increasing trade with Pakistan"
},
{
"idx": 8,
"label": 1,
"text": "Persuading Pakistan to use its influence over the Taliban by dangling before him the possibility of a presidential visit in March as a reward for Pakistani cooperation"
}
],
"idx": 0,
"question": "What did the high-level effort to persuade Pakistan include?"
},
...
],
"text": "While this process moved along, diplomacy ..."
}
ReCoRD
- train split: 65709 examples
- validation split: 7481 examples
- test split: 7484 examples
An example of 'train' looks as follows.
{
"source": "Daily mail",
"passage": {
"entities": [
{
"end": 96,
"start": 86
},
{
"end": 183,
"start": 178
},
{
"end": 206,
"start": 197
},
{
"end": 361,
"start": 357
},
{
"end": 539,
"start": 535
},
{
"end": 636,
"start": 627
},
{
"end": 681,
"start": 672
}
],
"text": "The harrowing stories of women and children locked up for so-called 'moral crimes' in Afghanistan's notorious female prison have been revealed after cameras were allowed inside. Mariam has been in Badam Bagh prison for three months after she shot a man who just raped her at gunpoint and then turned the weapon on herself - but she has yet to been charged. Nuria has eight months left to serve of her sentence for trying to divorce her husband. She gave birth in prison to her son and they share a cell together. Scroll down for video Nuria was jailed for trying to divorce her husband. Her son is one of 62 children living at Badam Bagh prison\n@highlight\nMost of the 202 Badam Bagh inmates are jailed for so-called 'moral crimes'\n@highlight\nCrimes include leaving their husbands or refusing an arrange marriage\n@highlight\n62 children live there and share cells with their mothers and five others"
},
"qas": [
{
"answers": [
{
"end": 539,
"start": 535,
"text": "Nuria"
}
],
"idx": 0,
"query": "The baby she gave birth to is her husbands and he has even offered to have the courts set her free if she returns, but @placeholder has refused."
}
],
"idx": 0
}
RTE
- train split: 2490 examples
- validation split: 277 examples
- test split: 3000 examples
An example of 'train' looks as follows.
{
"premise": "No Weapons of Mass Destruction Found in Iraq Yet.",
"hypothesis": "Weapons of Mass Destruction Found in Iraq.",
"label": "not_entailment",
"idx": 0
}
WiC
- train split: 5428 examples
- validation split: 638 examples
- test split: 1400 examples
An example of 'train' looks as follows.
{
"word": "place",
"sentence1": "Do you want to come over to my place later?",
"sentence2": "A political system with no place for the less prominent groups.",
"idx": 0,
"label": false,
"start1": 31,
"start2": 27,
"end1": 36,
"end2": 32,
"version": 1.1
}
WSC
- train split: 554 examples
- validation split: 104 examples
- test split: 146 examples
An example of 'train' looks as follows.
{
"text": "Mark told Pete many lies about himself, which Pete included in his book. He should have been more skeptical.",
"target": {
"span1_index": 0,
"span1_text": "Mark",
"span2_index": 13,
"span2_text": "He"
},
"idx": 0,
"label": false
}
Usage
You can load any of the datasets using the Huggingface datasets library. For example, to load the BoolQ dataset, run:
from datasets import load_dataset
# Load the BoolQ dataset from the SuperGLUE benchmark
dataset = load_dataset("Hyukkyu/superglue", "boolq")
# Access train, validation, and test splits
train_split = dataset["train"]
validation_split = dataset["validation"]
test_split = dataset["test"]
print(train_split)
Replace "boolq" with the desired configuration name (e.g., "cb", "copa", "multirc", etc.) to load other datasets.
Data Preprocessing
- Schema Consistency: A recursive procedure was used to infer the schema from the train and validation splits and fill in missing keys in the test split with dummy values. This ensures that all splits have the same features, preventing issues during model training or evaluation.
- Type-Aware Dummy Values: Dummy values are inserted based on the expected type. For instance, missing boolean fields are filled with False, integer fields with -1, float fields with -1.0, and string fields with an empty string.
Citation
@article{wang2019superglue,
title={Superglue: A stickier benchmark for general-purpose language understanding systems},
author={Wang, Alex and Pruksachatkun, Yada and Nangia, Nikita and Singh, Amanpreet and Michael, Julian and Hill, Felix and Levy, Omer and Bowman, Samuel},
journal={Advances in neural information processing systems},
volume={32},
year={2019}
}