Add "Adversarial GLUE" dataset to datasets library (#3849)
Browse files* update link in wiki_bio dataset
* run linter and update dummy data
* fix markdown so that test passes (even though I didnt break it)
* init adversarial glue in hf datasets
* format and add card
* update label computation
* remove backslash
* remove print statement
* Update datasets/adv_glue/README.md
Co-authored-by: Quentin Lhoest <42851186+lhoestq@users.noreply.github.com>
* Update datasets/adv_glue/README.md
Co-authored-by: Quentin Lhoest <42851186+lhoestq@users.noreply.github.com>
* flesh out README.md
* Update datasets/adv_glue/README.md
advglue
Co-authored-by: Quentin Lhoest <42851186+lhoestq@users.noreply.github.com>
* Update datasets/adv_glue/README.md
Co-authored-by: Quentin Lhoest <42851186+lhoestq@users.noreply.github.com>
* update tags and fields
* update tags
* minor changes to trigger the CI
Co-authored-by: Quentin Lhoest <42851186+lhoestq@users.noreply.github.com>
Commit from https://github.com/huggingface/datasets/commit/dfa8f056866dfa9efe32d6e90a8493546dac218d
- README.md +200 -0
- adv_glue.py +330 -0
- dataset_infos.json +1 -0
- dummy/adv_mnli/1.0.0/dummy_data.zip +3 -0
- dummy/adv_mnli_mismatched/1.0.0/dummy_data.zip +3 -0
- dummy/adv_qnli/1.0.0/dummy_data.zip +3 -0
- dummy/adv_qqp/1.0.0/dummy_data.zip +3 -0
- dummy/adv_rte/1.0.0/dummy_data.zip +3 -0
- dummy/adv_sst2/1.0.0/dummy_data.zip +3 -0
@@ -0,0 +1,200 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
annotations_creators:
|
3 |
+
- unknown
|
4 |
+
language_creators:
|
5 |
+
- machine-generated
|
6 |
+
languages:
|
7 |
+
- en
|
8 |
+
licenses:
|
9 |
+
- cc-by-4-0
|
10 |
+
multilinguality:
|
11 |
+
- monolingual
|
12 |
+
pretty_name: Adversarial GLUE
|
13 |
+
size_categories:
|
14 |
+
- n<1K
|
15 |
+
source_datasets:
|
16 |
+
- extended|glue
|
17 |
+
task_categories:
|
18 |
+
adv_mnli:
|
19 |
+
- text-classification
|
20 |
+
adv_mnli_mismatched:
|
21 |
+
- text-classification
|
22 |
+
adv_qnli:
|
23 |
+
- text-classification
|
24 |
+
adv_qqp:
|
25 |
+
- text-classification
|
26 |
+
adv_rte:
|
27 |
+
- text-classification
|
28 |
+
adv_sst2:
|
29 |
+
- text-classification
|
30 |
+
task_ids:
|
31 |
+
adv_mnli:
|
32 |
+
- natural-language-inference
|
33 |
+
adv_mnli_mismatched:
|
34 |
+
- natural-language-inference
|
35 |
+
adv_qnli:
|
36 |
+
- text-classification-other-qa-nli
|
37 |
+
adv_qqp:
|
38 |
+
- text-classification-other-paraphrase-identification
|
39 |
+
adv_rte:
|
40 |
+
- natural-language-inference
|
41 |
+
adv_sst2:
|
42 |
+
- sentiment-classification
|
43 |
+
---
|
44 |
+
|
45 |
+
# Dataset Card for Adversarial GLUE
|
46 |
+
|
47 |
+
## Table of Contents
|
48 |
+
- [Table of Contents](#table-of-contents)
|
49 |
+
- [Dataset Description](#dataset-description)
|
50 |
+
- [Dataset Summary](#dataset-summary)
|
51 |
+
- [Supported Tasks and Leaderboards](#supported-tasks-and-leaderboards)
|
52 |
+
- [Languages](#languages)
|
53 |
+
- [Dataset Structure](#dataset-structure)
|
54 |
+
- [Data Instances](#data-instances)
|
55 |
+
- [Data Fields](#data-fields)
|
56 |
+
- [Data Splits](#data-splits)
|
57 |
+
- [Dataset Creation](#dataset-creation)
|
58 |
+
- [Curation Rationale](#curation-rationale)
|
59 |
+
- [Source Data](#source-data)
|
60 |
+
- [Annotations](#annotations)
|
61 |
+
- [Personal and Sensitive Information](#personal-and-sensitive-information)
|
62 |
+
- [Considerations for Using the Data](#considerations-for-using-the-data)
|
63 |
+
- [Social Impact of Dataset](#social-impact-of-dataset)
|
64 |
+
- [Discussion of Biases](#discussion-of-biases)
|
65 |
+
- [Other Known Limitations](#other-known-limitations)
|
66 |
+
- [Additional Information](#additional-information)
|
67 |
+
- [Dataset Curators](#dataset-curators)
|
68 |
+
- [Licensing Information](#licensing-information)
|
69 |
+
- [Citation Information](#citation-information)
|
70 |
+
- [Contributions](#contributions)
|
71 |
+
|
72 |
+
|
73 |
+
## Dataset Description
|
74 |
+
|
75 |
+
- **Homepage:** https://adversarialglue.github.io/
|
76 |
+
- **Repository:**
|
77 |
+
- **Paper:** [arXiv](https://arxiv.org/pdf/2111.02840.pdf)
|
78 |
+
- **Leaderboard:**
|
79 |
+
- **Point of Contact:**
|
80 |
+
Adversarial GLUE Benchmark (AdvGLUE) is a comprehensive robustness evaluation benchmark that focuses on the adversarial robustness evaluation of language models. It covers five natural language understanding tasks from the famous GLUE tasks and is an adversarial version of GLUE benchmark.
|
81 |
+
|
82 |
+
AdvGLUE considers textual adversarial attacks from different perspectives and hierarchies, including word-level transformations, sentence-level manipulations, and human-written adversarial examples, which provide comprehensive coverage of various adversarial linguistic phenomena.
|
83 |
+
|
84 |
+
### Supported Tasks and Leaderboards
|
85 |
+
|
86 |
+
Leaderboard available on the homepage: [https://adversarialglue.github.io/](https://adversarialglue.github.io/).
|
87 |
+
|
88 |
+
### Languages
|
89 |
+
|
90 |
+
AdvGLUE deviates from the GLUE dataset, which has a base language of English.
|
91 |
+
|
92 |
+
## Dataset Structure
|
93 |
+
|
94 |
+
We show detailed information for up to 5 configurations of the dataset.
|
95 |
+
|
96 |
+
### Data Instances
|
97 |
+
|
98 |
+
#### default
|
99 |
+
|
100 |
+
- **Size of downloaded dataset files:** 198 KB
|
101 |
+
- **Example**:
|
102 |
+
```python
|
103 |
+
>>> datasets.load_dataset('adv_glue', 'adv_sst2')['validation'][0]
|
104 |
+
{'sentence': "it 's an uneven treat that bores fun at the democratic exercise while also examining its significance for those who take part .", 'label': 1, 'idx': 0}
|
105 |
+
```
|
106 |
+
|
107 |
+
### Data Fields
|
108 |
+
|
109 |
+
The data fields are the same as in the GLUE dataset, which differ by task.
|
110 |
+
|
111 |
+
|
112 |
+
The data fields are the same among all splits.
|
113 |
+
|
114 |
+
#### adv_mnli
|
115 |
+
- `premise`: a `string` feature.
|
116 |
+
- `hypothesis`: a `string` feature.
|
117 |
+
- `label`: a classification label, with possible values including `entailment` (0), `neutral` (1), `contradiction` (2).
|
118 |
+
- `idx`: a `int32` feature.
|
119 |
+
|
120 |
+
#### adv_mnli_matched
|
121 |
+
- `premise`: a `string` feature.
|
122 |
+
- `hypothesis`: a `string` feature.
|
123 |
+
- `label`: a classification label, with possible values including `entailment` (0), `neutral` (1), `contradiction` (2).
|
124 |
+
- `idx`: a `int32` feature.
|
125 |
+
|
126 |
+
#### adv_mnli_mismatched
|
127 |
+
- `premise`: a `string` feature.
|
128 |
+
- `hypothesis`: a `string` feature.
|
129 |
+
- `label`: a classification label, with possible values including `entailment` (0), `neutral` (1), `contradiction` (2).
|
130 |
+
- `idx`: a `int32` feature.
|
131 |
+
|
132 |
+
#### adv_qnli
|
133 |
+
|
134 |
+
[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
|
135 |
+
|
136 |
+
#### adv_qqp
|
137 |
+
|
138 |
+
[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
|
139 |
+
|
140 |
+
#### adv_rte
|
141 |
+
|
142 |
+
[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
|
143 |
+
|
144 |
+
#### adv_sst2
|
145 |
+
|
146 |
+
[More Information Needed](https://github.com/huggingface/datasets/blob/master/CONTRIBUTING.md#how-to-contribute-to-the-dataset-cards)
|
147 |
+
|
148 |
+
### Data Splits
|
149 |
+
|
150 |
+
Adversarial GLUE provides only a 'dev' split.
|
151 |
+
|
152 |
+
## Dataset Creation
|
153 |
+
|
154 |
+
### Curation Rationale
|
155 |
+
|
156 |
+
[More Information Needed]
|
157 |
+
|
158 |
+
### Source Data
|
159 |
+
|
160 |
+
[More Information Needed]
|
161 |
+
|
162 |
+
### Annotations
|
163 |
+
|
164 |
+
[More Information Needed]
|
165 |
+
|
166 |
+
### Personal and Sensitive Information
|
167 |
+
|
168 |
+
[More Information Needed]
|
169 |
+
|
170 |
+
## Considerations for Using the Data
|
171 |
+
|
172 |
+
### Social Impact of Dataset
|
173 |
+
|
174 |
+
[More Information Needed]
|
175 |
+
|
176 |
+
### Discussion of Biases
|
177 |
+
|
178 |
+
[More Information Needed]
|
179 |
+
|
180 |
+
### Other Known Limitations
|
181 |
+
|
182 |
+
[More Information Needed]
|
183 |
+
|
184 |
+
## Additional Information
|
185 |
+
|
186 |
+
### Citation Information
|
187 |
+
|
188 |
+
```bibtex
|
189 |
+
@article{Wang2021AdversarialGA,
|
190 |
+
title={Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models},
|
191 |
+
author={Boxin Wang and Chejian Xu and Shuohang Wang and Zhe Gan and Yu Cheng and Jianfeng Gao and Ahmed Hassan Awadallah and B. Li},
|
192 |
+
journal={ArXiv},
|
193 |
+
year={2021},
|
194 |
+
volume={abs/2111.02840}
|
195 |
+
}
|
196 |
+
```
|
197 |
+
|
198 |
+
### Contributions
|
199 |
+
|
200 |
+
Thanks to [@jxmorris12](https://github.com/jxmorris12) for adding this dataset.
|
@@ -0,0 +1,330 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""The Adversarial GLUE (AdvGLUE) benchmark.
|
2 |
+
Homepage: https://adversarialglue.github.io/
|
3 |
+
"""
|
4 |
+
import json
|
5 |
+
import os
|
6 |
+
import textwrap
|
7 |
+
|
8 |
+
import datasets
|
9 |
+
|
10 |
+
|
11 |
+
_ADV_GLUE_CITATION = """\
|
12 |
+
@article{Wang2021AdversarialGA,
|
13 |
+
title={Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models},
|
14 |
+
author={Boxin Wang and Chejian Xu and Shuohang Wang and Zhe Gan and Yu Cheng and Jianfeng Gao and Ahmed Hassan Awadallah and B. Li},
|
15 |
+
journal={ArXiv},
|
16 |
+
year={2021},
|
17 |
+
volume={abs/2111.02840}
|
18 |
+
}
|
19 |
+
"""
|
20 |
+
|
21 |
+
_ADV_GLUE_DESCRIPTION = """\
|
22 |
+
Adversarial GLUE Benchmark (AdvGLUE) is a comprehensive robustness evaluation benchmark
|
23 |
+
that focuses on the adversarial robustness evaluation of language models. It covers five
|
24 |
+
natural language understanding tasks from the famous GLUE tasks and is an adversarial
|
25 |
+
version of GLUE benchmark.
|
26 |
+
"""
|
27 |
+
|
28 |
+
_MNLI_BASE_KWARGS = dict(
|
29 |
+
text_features={
|
30 |
+
"premise": "premise",
|
31 |
+
"hypothesis": "hypothesis",
|
32 |
+
},
|
33 |
+
label_classes=["entailment", "neutral", "contradiction"],
|
34 |
+
label_column="label",
|
35 |
+
data_url="https://dl.fbaipublicfiles.com/glue/data/MNLI.zip",
|
36 |
+
data_dir="MNLI",
|
37 |
+
citation=textwrap.dedent(
|
38 |
+
"""\
|
39 |
+
@InProceedings{N18-1101,
|
40 |
+
author = "Williams, Adina
|
41 |
+
and Nangia, Nikita
|
42 |
+
and Bowman, Samuel",
|
43 |
+
title = "A Broad-Coverage Challenge Corpus for
|
44 |
+
Sentence Understanding through Inference",
|
45 |
+
booktitle = "Proceedings of the 2018 Conference of
|
46 |
+
the North American Chapter of the
|
47 |
+
Association for Computational Linguistics:
|
48 |
+
Human Language Technologies, Volume 1 (Long
|
49 |
+
Papers)",
|
50 |
+
year = "2018",
|
51 |
+
publisher = "Association for Computational Linguistics",
|
52 |
+
pages = "1112--1122",
|
53 |
+
location = "New Orleans, Louisiana",
|
54 |
+
url = "http://aclweb.org/anthology/N18-1101"
|
55 |
+
}
|
56 |
+
@article{bowman2015large,
|
57 |
+
title={A large annotated corpus for learning natural language inference},
|
58 |
+
author={Bowman, Samuel R and Angeli, Gabor and Potts, Christopher and Manning, Christopher D},
|
59 |
+
journal={arXiv preprint arXiv:1508.05326},
|
60 |
+
year={2015}
|
61 |
+
}"""
|
62 |
+
),
|
63 |
+
url="http://www.nyu.edu/projects/bowman/multinli/",
|
64 |
+
)
|
65 |
+
|
66 |
+
ADVGLUE_DEV_URL = "https://adversarialglue.github.io/dataset/dev.zip"
|
67 |
+
|
68 |
+
|
69 |
+
class AdvGlueConfig(datasets.BuilderConfig):
|
70 |
+
"""BuilderConfig for Adversarial GLUE."""
|
71 |
+
|
72 |
+
def __init__(
|
73 |
+
self,
|
74 |
+
text_features,
|
75 |
+
label_column,
|
76 |
+
data_url,
|
77 |
+
data_dir,
|
78 |
+
citation,
|
79 |
+
url,
|
80 |
+
label_classes=None,
|
81 |
+
process_label=lambda x: x,
|
82 |
+
**kwargs,
|
83 |
+
):
|
84 |
+
"""BuilderConfig for Adversarial GLUE.
|
85 |
+
|
86 |
+
Args:
|
87 |
+
text_features: `dict[string, string]`, map from the name of the feature
|
88 |
+
dict for each text field to the name of the column in the tsv file
|
89 |
+
label_column: `string`, name of the column in the tsv file corresponding
|
90 |
+
to the label
|
91 |
+
data_url: `string`, url to download the zip file from
|
92 |
+
data_dir: `string`, the path to the folder containing the tsv files in the
|
93 |
+
downloaded zip
|
94 |
+
citation: `string`, citation for the data set
|
95 |
+
url: `string`, url for information about the data set
|
96 |
+
label_classes: `list[string]`, the list of classes if the label is
|
97 |
+
categorical. If not provided, then the label will be of type
|
98 |
+
`datasets.Value('float32')`.
|
99 |
+
process_label: `Function[string, any]`, function taking in the raw value
|
100 |
+
of the label and processing it to the form required by the label feature
|
101 |
+
**kwargs: keyword arguments forwarded to super.
|
102 |
+
"""
|
103 |
+
super(AdvGlueConfig, self).__init__(version=datasets.Version("1.0.0", ""), **kwargs)
|
104 |
+
self.text_features = text_features
|
105 |
+
self.label_column = label_column
|
106 |
+
self.label_classes = label_classes
|
107 |
+
self.data_url = data_url
|
108 |
+
self.data_dir = data_dir
|
109 |
+
self.citation = citation
|
110 |
+
self.url = url
|
111 |
+
self.process_label = process_label
|
112 |
+
|
113 |
+
|
114 |
+
ADVGLUE_BUILDER_CONFIGS = [
|
115 |
+
AdvGlueConfig(
|
116 |
+
name="adv_sst2",
|
117 |
+
description=textwrap.dedent(
|
118 |
+
"""Adversarial version of SST-2.
|
119 |
+
The Stanford Sentiment Treebank consists of sentences from movie reviews and
|
120 |
+
human annotations of their sentiment. The task is to predict the sentiment of a
|
121 |
+
given sentence. We use the two-way (positive/negative) class split, and use only
|
122 |
+
sentence-level labels."""
|
123 |
+
),
|
124 |
+
text_features={"sentence": "sentence"},
|
125 |
+
label_classes=["negative", "positive"],
|
126 |
+
label_column="label",
|
127 |
+
data_url="https://dl.fbaipublicfiles.com/glue/data/SST-2.zip",
|
128 |
+
data_dir="SST-2",
|
129 |
+
citation=textwrap.dedent(
|
130 |
+
"""\
|
131 |
+
@inproceedings{socher2013recursive,
|
132 |
+
title={Recursive deep models for semantic compositionality over a sentiment treebank},
|
133 |
+
author={Socher, Richard and Perelygin, Alex and Wu, Jean and Chuang, Jason and Manning, Christopher D and Ng, Andrew and Potts, Christopher},
|
134 |
+
booktitle={Proceedings of the 2013 conference on empirical methods in natural language processing},
|
135 |
+
pages={1631--1642},
|
136 |
+
year={2013}
|
137 |
+
}"""
|
138 |
+
),
|
139 |
+
url="https://datasets.stanford.edu/sentiment/index.html",
|
140 |
+
),
|
141 |
+
AdvGlueConfig(
|
142 |
+
name="adv_qqp",
|
143 |
+
description=textwrap.dedent(
|
144 |
+
"""Adversarial version of QQP.
|
145 |
+
The Quora Question Pairs2 dataset is a collection of question pairs from the
|
146 |
+
community question-answering website Quora. The task is to determine whether a
|
147 |
+
pair of questions are semantically equivalent."""
|
148 |
+
),
|
149 |
+
text_features={
|
150 |
+
"question1": "question1",
|
151 |
+
"question2": "question2",
|
152 |
+
},
|
153 |
+
label_classes=["not_duplicate", "duplicate"],
|
154 |
+
label_column="label",
|
155 |
+
data_url="https://dl.fbaipublicfiles.com/glue/data/QQP-clean.zip",
|
156 |
+
data_dir="QQP",
|
157 |
+
citation=textwrap.dedent(
|
158 |
+
"""\
|
159 |
+
@online{WinNT,
|
160 |
+
author = {Iyer, Shankar and Dandekar, Nikhil and Csernai, Kornel},
|
161 |
+
title = {First Quora Dataset Release: Question Pairs},
|
162 |
+
year = {2017},
|
163 |
+
url = {https://data.quora.com/First-Quora-Dataset-Release-Question-Pairs},
|
164 |
+
urldate = {2019-04-03}
|
165 |
+
}"""
|
166 |
+
),
|
167 |
+
url="https://data.quora.com/First-Quora-Dataset-Release-Question-Pairs",
|
168 |
+
),
|
169 |
+
AdvGlueConfig(
|
170 |
+
name="adv_mnli",
|
171 |
+
description=textwrap.dedent(
|
172 |
+
"""Adversarial version of MNLI.
|
173 |
+
The Multi-Genre Natural Language Inference Corpus is a crowdsourced
|
174 |
+
collection of sentence pairs with textual entailment annotations. Given a premise sentence
|
175 |
+
and a hypothesis sentence, the task is to predict whether the premise entails the hypothesis
|
176 |
+
(entailment), contradicts the hypothesis (contradiction), or neither (neutral). The premise sentences are
|
177 |
+
gathered from ten different sources, including transcribed speech, fiction, and government reports.
|
178 |
+
We use the standard test set, for which we obtained private labels from the authors, and evaluate
|
179 |
+
on both the matched (in-domain) and mismatched (cross-domain) section. We also use and recommend
|
180 |
+
the SNLI corpus as 550k examples of auxiliary training data."""
|
181 |
+
),
|
182 |
+
**_MNLI_BASE_KWARGS,
|
183 |
+
),
|
184 |
+
AdvGlueConfig(
|
185 |
+
name="adv_mnli_mismatched",
|
186 |
+
description=textwrap.dedent(
|
187 |
+
"""Adversarial version of MNLI-mismatched.
|
188 |
+
The mismatched validation and test splits from MNLI.
|
189 |
+
See the "mnli" BuilderConfig for additional information."""
|
190 |
+
),
|
191 |
+
**_MNLI_BASE_KWARGS,
|
192 |
+
),
|
193 |
+
AdvGlueConfig(
|
194 |
+
name="adv_qnli",
|
195 |
+
description=textwrap.dedent(
|
196 |
+
"""Adversarial version of QNLI.
|
197 |
+
The Stanford Question Answering Dataset is a question-answering
|
198 |
+
dataset consisting of question-paragraph pairs, where one of the sentences in the paragraph (drawn
|
199 |
+
from Wikipedia) contains the answer to the corresponding question (written by an annotator). We
|
200 |
+
convert the task into sentence pair classification by forming a pair between each question and each
|
201 |
+
sentence in the corresponding context, and filtering out pairs with low lexical overlap between the
|
202 |
+
question and the context sentence. The task is to determine whether the context sentence contains
|
203 |
+
the answer to the question. This modified version of the original task removes the requirement that
|
204 |
+
the model select the exact answer, but also removes the simplifying assumptions that the answer
|
205 |
+
is always present in the input and that lexical overlap is a reliable cue."""
|
206 |
+
), # pylint: disable=line-too-long
|
207 |
+
text_features={
|
208 |
+
"question": "question",
|
209 |
+
"sentence": "sentence",
|
210 |
+
},
|
211 |
+
label_classes=["entailment", "not_entailment"],
|
212 |
+
label_column="label",
|
213 |
+
data_url="https://dl.fbaipublicfiles.com/glue/data/QNLIv2.zip",
|
214 |
+
data_dir="QNLI",
|
215 |
+
citation=textwrap.dedent(
|
216 |
+
"""\
|
217 |
+
@article{rajpurkar2016squad,
|
218 |
+
title={Squad: 100,000+ questions for machine comprehension of text},
|
219 |
+
author={Rajpurkar, Pranav and Zhang, Jian and Lopyrev, Konstantin and Liang, Percy},
|
220 |
+
journal={arXiv preprint arXiv:1606.05250},
|
221 |
+
year={2016}
|
222 |
+
}"""
|
223 |
+
),
|
224 |
+
url="https://rajpurkar.github.io/SQuAD-explorer/",
|
225 |
+
),
|
226 |
+
AdvGlueConfig(
|
227 |
+
name="adv_rte",
|
228 |
+
description=textwrap.dedent(
|
229 |
+
"""Adversarial version of RTE.
|
230 |
+
The Recognizing Textual Entailment (RTE) datasets come from a series of annual textual
|
231 |
+
entailment challenges. We combine the data from RTE1 (Dagan et al., 2006), RTE2 (Bar Haim
|
232 |
+
et al., 2006), RTE3 (Giampiccolo et al., 2007), and RTE5 (Bentivogli et al., 2009).4 Examples are
|
233 |
+
constructed based on news and Wikipedia text. We convert all datasets to a two-class split, where
|
234 |
+
for three-class datasets we collapse neutral and contradiction into not entailment, for consistency."""
|
235 |
+
), # pylint: disable=line-too-long
|
236 |
+
text_features={
|
237 |
+
"sentence1": "sentence1",
|
238 |
+
"sentence2": "sentence2",
|
239 |
+
},
|
240 |
+
label_classes=["entailment", "not_entailment"],
|
241 |
+
label_column="label",
|
242 |
+
data_url="https://dl.fbaipublicfiles.com/glue/data/RTE.zip",
|
243 |
+
data_dir="RTE",
|
244 |
+
citation=textwrap.dedent(
|
245 |
+
"""\
|
246 |
+
@inproceedings{dagan2005pascal,
|
247 |
+
title={The PASCAL recognising textual entailment challenge},
|
248 |
+
author={Dagan, Ido and Glickman, Oren and Magnini, Bernardo},
|
249 |
+
booktitle={Machine Learning Challenges Workshop},
|
250 |
+
pages={177--190},
|
251 |
+
year={2005},
|
252 |
+
organization={Springer}
|
253 |
+
}
|
254 |
+
@inproceedings{bar2006second,
|
255 |
+
title={The second pascal recognising textual entailment challenge},
|
256 |
+
author={Bar-Haim, Roy and Dagan, Ido and Dolan, Bill and Ferro, Lisa and Giampiccolo, Danilo and Magnini, Bernardo and Szpektor, Idan},
|
257 |
+
booktitle={Proceedings of the second PASCAL challenges workshop on recognising textual entailment},
|
258 |
+
volume={6},
|
259 |
+
number={1},
|
260 |
+
pages={6--4},
|
261 |
+
year={2006},
|
262 |
+
organization={Venice}
|
263 |
+
}
|
264 |
+
@inproceedings{giampiccolo2007third,
|
265 |
+
title={The third pascal recognizing textual entailment challenge},
|
266 |
+
author={Giampiccolo, Danilo and Magnini, Bernardo and Dagan, Ido and Dolan, Bill},
|
267 |
+
booktitle={Proceedings of the ACL-PASCAL workshop on textual entailment and paraphrasing},
|
268 |
+
pages={1--9},
|
269 |
+
year={2007},
|
270 |
+
organization={Association for Computational Linguistics}
|
271 |
+
}
|
272 |
+
@inproceedings{bentivogli2009fifth,
|
273 |
+
title={The Fifth PASCAL Recognizing Textual Entailment Challenge.},
|
274 |
+
author={Bentivogli, Luisa and Clark, Peter and Dagan, Ido and Giampiccolo, Danilo},
|
275 |
+
booktitle={TAC},
|
276 |
+
year={2009}
|
277 |
+
}"""
|
278 |
+
),
|
279 |
+
url="https://aclweb.org/aclwiki/Recognizing_Textual_Entailment",
|
280 |
+
),
|
281 |
+
]
|
282 |
+
|
283 |
+
|
284 |
+
class AdvGlue(datasets.GeneratorBasedBuilder):
|
285 |
+
"""The General Language Understanding Evaluation (GLUE) benchmark."""
|
286 |
+
|
287 |
+
DATASETS = ["adv_sst2", "adv_qqp", "adv_mnli", "adv_mnli_mismatched", "adv_qnli", "adv_rte"]
|
288 |
+
BUILDER_CONFIGS = ADVGLUE_BUILDER_CONFIGS
|
289 |
+
|
290 |
+
def _info(self):
|
291 |
+
features = {text_feature: datasets.Value("string") for text_feature in self.config.text_features.keys()}
|
292 |
+
if self.config.label_classes:
|
293 |
+
features["label"] = datasets.features.ClassLabel(names=self.config.label_classes)
|
294 |
+
else:
|
295 |
+
features["label"] = datasets.Value("float32")
|
296 |
+
features["idx"] = datasets.Value("int32")
|
297 |
+
return datasets.DatasetInfo(
|
298 |
+
description=_ADV_GLUE_DESCRIPTION,
|
299 |
+
features=datasets.Features(features),
|
300 |
+
homepage="https://adversarialglue.github.io/",
|
301 |
+
citation=_ADV_GLUE_CITATION,
|
302 |
+
)
|
303 |
+
|
304 |
+
def _split_generators(self, dl_manager):
|
305 |
+
assert self.config.name in AdvGlue.DATASETS
|
306 |
+
data_dir = dl_manager.download_and_extract(ADVGLUE_DEV_URL)
|
307 |
+
data_file = os.path.join(data_dir, "dev", "dev.json")
|
308 |
+
return [
|
309 |
+
datasets.SplitGenerator(
|
310 |
+
name=datasets.Split.VALIDATION,
|
311 |
+
gen_kwargs={
|
312 |
+
"data_file": data_file,
|
313 |
+
},
|
314 |
+
)
|
315 |
+
]
|
316 |
+
|
317 |
+
def _generate_examples(self, data_file):
|
318 |
+
# We name splits 'adv_sst2' instead of 'sst2' so as not to be confused
|
319 |
+
# with the original SST-2. Here they're named like 'sst2' so we have to
|
320 |
+
# remove the 'adv_' prefix.
|
321 |
+
config_key = self.config.name.replace("adv_", "")
|
322 |
+
if config_key == "mnli_mismatched":
|
323 |
+
# and they name this split differently.
|
324 |
+
config_key = "mnli-mm"
|
325 |
+
data = json.loads(open(data_file).read())
|
326 |
+
for row in data[config_key]:
|
327 |
+
example = {feat: row[col] for feat, col in self.config.text_features.items()}
|
328 |
+
example["label"] = self.config.process_label(row[self.config.label_column])
|
329 |
+
example["idx"] = row["idx"]
|
330 |
+
yield example["idx"], example
|
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"adv_sst2": {"description": "Adversarial GLUE Benchmark (AdvGLUE) is a comprehensive robustness evaluation benchmark\nthat focuses on the adversarial robustness evaluation of language models. It covers five\nnatural language understanding tasks from the famous GLUE tasks and is an adversarial\nversion of GLUE benchmark.\n", "citation": "@article{Wang2021AdversarialGA,\n title={Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models},\n author={Boxin Wang and Chejian Xu and Shuohang Wang and Zhe Gan and Yu Cheng and Jianfeng Gao and Ahmed Hassan Awadallah and B. Li},\n journal={ArXiv},\n year={2021},\n volume={abs/2111.02840}\n}\n", "homepage": "https://adversarialglue.github.io/", "license": "", "features": {"sentence": {"dtype": "string", "id": null, "_type": "Value"}, "idx": {"dtype": "int32", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "adv_glue", "config_name": "adv_sst2", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 15407, "num_examples": 148, "dataset_name": "adv_glue"}}, "download_checksums": {"https://adversarialglue.github.io/dataset/dev.zip": {"num_bytes": 40662, "checksum": null}}, "download_size": 40662, "post_processing_size": null, "dataset_size": 15407, "size_in_bytes": 56069}, "adv_qqp": {"description": "Adversarial GLUE Benchmark (AdvGLUE) is a comprehensive robustness evaluation benchmark\nthat focuses on the adversarial robustness evaluation of language models. It covers five\nnatural language understanding tasks from the famous GLUE tasks and is an adversarial\nversion of GLUE benchmark.\n", "citation": "@article{Wang2021AdversarialGA,\n title={Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models},\n author={Boxin Wang and Chejian Xu and Shuohang Wang and Zhe Gan and Yu Cheng and Jianfeng Gao and Ahmed Hassan Awadallah and B. Li},\n journal={ArXiv},\n year={2021},\n volume={abs/2111.02840}\n}\n", "homepage": "https://adversarialglue.github.io/", "license": "", "features": {"question1": {"dtype": "string", "id": null, "_type": "Value"}, "question2": {"dtype": "string", "id": null, "_type": "Value"}, "idx": {"dtype": "int32", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "adv_glue", "config_name": "adv_qqp", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 9294, "num_examples": 78, "dataset_name": "adv_glue"}}, "download_checksums": {"https://adversarialglue.github.io/dataset/dev.zip": {"num_bytes": 40662, "checksum": null}}, "download_size": 40662, "post_processing_size": null, "dataset_size": 9294, "size_in_bytes": 49956}, "adv_mnli": {"description": "Adversarial GLUE Benchmark (AdvGLUE) is a comprehensive robustness evaluation benchmark\nthat focuses on the adversarial robustness evaluation of language models. It covers five\nnatural language understanding tasks from the famous GLUE tasks and is an adversarial\nversion of GLUE benchmark.\n", "citation": "@article{Wang2021AdversarialGA,\n title={Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models},\n author={Boxin Wang and Chejian Xu and Shuohang Wang and Zhe Gan and Yu Cheng and Jianfeng Gao and Ahmed Hassan Awadallah and B. Li},\n journal={ArXiv},\n year={2021},\n volume={abs/2111.02840}\n}\n", "homepage": "https://adversarialglue.github.io/", "license": "", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "idx": {"dtype": "int32", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "adv_glue", "config_name": "adv_mnli", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 22760, "num_examples": 121, "dataset_name": "adv_glue"}}, "download_checksums": {"https://adversarialglue.github.io/dataset/dev.zip": {"num_bytes": 40662, "checksum": null}}, "download_size": 40662, "post_processing_size": null, "dataset_size": 22760, "size_in_bytes": 63422}, "adv_mnli_mismatched": {"description": "Adversarial GLUE Benchmark (AdvGLUE) is a comprehensive robustness evaluation benchmark\nthat focuses on the adversarial robustness evaluation of language models. It covers five\nnatural language understanding tasks from the famous GLUE tasks and is an adversarial\nversion of GLUE benchmark.\n", "citation": "@article{Wang2021AdversarialGA,\n title={Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models},\n author={Boxin Wang and Chejian Xu and Shuohang Wang and Zhe Gan and Yu Cheng and Jianfeng Gao and Ahmed Hassan Awadallah and B. Li},\n journal={ArXiv},\n year={2021},\n volume={abs/2111.02840}\n}\n", "homepage": "https://adversarialglue.github.io/", "license": "", "features": {"premise": {"dtype": "string", "id": null, "_type": "Value"}, "hypothesis": {"dtype": "string", "id": null, "_type": "Value"}, "idx": {"dtype": "int32", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "adv_glue", "config_name": "adv_mnli_mismatched", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 39678, "num_examples": 162, "dataset_name": "adv_glue"}}, "download_checksums": {"https://adversarialglue.github.io/dataset/dev.zip": {"num_bytes": 40662, "checksum": null}}, "download_size": 40662, "post_processing_size": null, "dataset_size": 39678, "size_in_bytes": 80340}, "adv_qnli": {"description": "Adversarial GLUE Benchmark (AdvGLUE) is a comprehensive robustness evaluation benchmark\nthat focuses on the adversarial robustness evaluation of language models. It covers five\nnatural language understanding tasks from the famous GLUE tasks and is an adversarial\nversion of GLUE benchmark.\n", "citation": "@article{Wang2021AdversarialGA,\n title={Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models},\n author={Boxin Wang and Chejian Xu and Shuohang Wang and Zhe Gan and Yu Cheng and Jianfeng Gao and Ahmed Hassan Awadallah and B. Li},\n journal={ArXiv},\n year={2021},\n volume={abs/2111.02840}\n}\n", "homepage": "https://adversarialglue.github.io/", "license": "", "features": {"question": {"dtype": "string", "id": null, "_type": "Value"}, "sentence": {"dtype": "string", "id": null, "_type": "Value"}, "idx": {"dtype": "int32", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "adv_glue", "config_name": "adv_qnli", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 33685, "num_examples": 148, "dataset_name": "adv_glue"}}, "download_checksums": {"https://adversarialglue.github.io/dataset/dev.zip": {"num_bytes": 40662, "checksum": null}}, "download_size": 40662, "post_processing_size": null, "dataset_size": 33685, "size_in_bytes": 74347}, "adv_rte": {"description": "Adversarial GLUE Benchmark (AdvGLUE) is a comprehensive robustness evaluation benchmark\nthat focuses on the adversarial robustness evaluation of language models. It covers five\nnatural language understanding tasks from the famous GLUE tasks and is an adversarial\nversion of GLUE benchmark.\n", "citation": "@article{Wang2021AdversarialGA,\n title={Adversarial GLUE: A Multi-Task Benchmark for Robustness Evaluation of Language Models},\n author={Boxin Wang and Chejian Xu and Shuohang Wang and Zhe Gan and Yu Cheng and Jianfeng Gao and Ahmed Hassan Awadallah and B. Li},\n journal={ArXiv},\n year={2021},\n volume={abs/2111.02840}\n}\n", "homepage": "https://adversarialglue.github.io/", "license": "", "features": {"sentence1": {"dtype": "string", "id": null, "_type": "Value"}, "sentence2": {"dtype": "string", "id": null, "_type": "Value"}, "idx": {"dtype": "int32", "id": null, "_type": "Value"}}, "post_processed": null, "supervised_keys": null, "task_templates": null, "builder_name": "adv_glue", "config_name": "adv_rte", "version": {"version_str": "1.0.0", "description": "", "major": 1, "minor": 0, "patch": 0}, "splits": {"validation": {"name": "validation", "num_bytes": 25342, "num_examples": 81, "dataset_name": "adv_glue"}}, "download_checksums": {"https://adversarialglue.github.io/dataset/dev.zip": {"num_bytes": 40662, "checksum": null}}, "download_size": 40662, "post_processing_size": null, "dataset_size": 25342, "size_in_bytes": 66004}}
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5f8f8646ef0f99e2c6a1879bca7195c262e353871814a70f0239ea35d97db907
|
3 |
+
size 3252
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5f8f8646ef0f99e2c6a1879bca7195c262e353871814a70f0239ea35d97db907
|
3 |
+
size 3252
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5f8f8646ef0f99e2c6a1879bca7195c262e353871814a70f0239ea35d97db907
|
3 |
+
size 3252
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5f8f8646ef0f99e2c6a1879bca7195c262e353871814a70f0239ea35d97db907
|
3 |
+
size 3252
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:c26ec2d43a0e34c979c9bc93fde795032948c2b642ea2075ad54fbd98c666a51
|
3 |
+
size 3252
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:5f8f8646ef0f99e2c6a1879bca7195c262e353871814a70f0239ea35d97db907
|
3 |
+
size 3252
|