Datasets:

Modalities:
Text
Formats:
parquet
Languages:
English
ArXiv:
Libraries:
Datasets
pandas
License:
albertvillanova HF staff commited on
Commit
74e4b2c
1 Parent(s): 908855b

Add dbidaf data files

Browse files
README.md CHANGED
@@ -113,16 +113,16 @@ dataset_info:
113
  dtype: string
114
  splits:
115
  - name: train
116
- num_bytes: 9282518
117
  num_examples: 10000
118
  - name: validation
119
- num_bytes: 917943
120
  num_examples: 1000
121
  - name: test
122
- num_bytes: 947111
123
  num_examples: 1000
124
- download_size: 9018914
125
- dataset_size: 11147572
126
  - config_name: droberta
127
  features:
128
  - name: id
@@ -166,6 +166,14 @@ configs:
166
  path: adversarialQA/validation-*
167
  - split: test
168
  path: adversarialQA/test-*
 
 
 
 
 
 
 
 
169
  train-eval-index:
170
  - config: adversarialQA
171
  task: question-answering
 
113
  dtype: string
114
  splits:
115
  - name: train
116
+ num_bytes: 9282482
117
  num_examples: 10000
118
  - name: validation
119
+ num_bytes: 917907
120
  num_examples: 1000
121
  - name: test
122
+ num_bytes: 946947
123
  num_examples: 1000
124
+ download_size: 2721341
125
+ dataset_size: 11147336
126
  - config_name: droberta
127
  features:
128
  - name: id
 
166
  path: adversarialQA/validation-*
167
  - split: test
168
  path: adversarialQA/test-*
169
+ - config_name: dbidaf
170
+ data_files:
171
+ - split: train
172
+ path: dbidaf/train-*
173
+ - split: validation
174
+ path: dbidaf/validation-*
175
+ - split: test
176
+ path: dbidaf/test-*
177
  train-eval-index:
178
  - config: adversarialQA
179
  task: question-answering
dataset_infos.json CHANGED
@@ -80,67 +80,55 @@
80
  },
81
  "dbidaf": {
82
  "description": "AdversarialQA is a Reading Comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles using an adversarial model-in-the-loop.\nWe use three different models; BiDAF (Seo et al., 2016), BERT-Large (Devlin et al., 2018), and RoBERTa-Large (Liu et al., 2019) in the annotation loop and construct three datasets; D(BiDAF), D(BERT), and D(RoBERTa), each with 10,000 training examples, 1,000 validation, and 1,000 test examples.\nThe adversarial human annotation paradigm ensures that these datasets consist of questions that current state-of-the-art models (at least the ones used as adversaries in the annotation loop) find challenging.\n",
83
- "citation": "@article{bartolo2020beat,\n author = {Bartolo, Max and Roberts, Alastair and Welbl, Johannes and Riedel, Sebastian and Stenetorp, Pontus},\n title = {Beat the AI: Investigating Adversarial Human Annotation for Reading Comprehension},\n journal = {Transactions of the Association for Computational Linguistics},\n volume = {8},\n number = {},\n pages = {662-678},\n year = {2020},\n doi = {10.1162/tacl\\_a\\_00338},\n URL = { https://doi.org/10.1162/tacl_a_00338 },\n eprint = { https://doi.org/10.1162/tacl_a_00338 },\n abstract = { Innovations in annotation methodology have been a catalyst for Reading Comprehension (RC) datasets and models. One recent trend to challenge current RC models is to involve a model in the annotation process: Humans create questions adversarially, such that the model fails to answer them correctly. In this work we investigate this annotation methodology and apply it in three different settings, collecting a total of 36,000 samples with progressively stronger models in the annotation loop. This allows us to explore questions such as the reproducibility of the adversarial effect, transfer from data collected with varying model-in-the-loop strengths, and generalization to data collected without a model. We find that training on adversarially collected samples leads to strong generalization to non-adversarially collected datasets, yet with progressive performance deterioration with increasingly stronger models-in-the-loop. Furthermore, we find that stronger models can still learn from datasets collected with substantially weaker models-in-the-loop. When trained on data collected with a BiDAF model in the loop, RoBERTa achieves 39.9F1 on questions that it cannot answer when trained on SQuAD\u2014only marginally lower than when trained on data collected using RoBERTa itself (41.0F1). }\n}\n",
84
  "homepage": "https://adversarialqa.github.io/",
85
  "license": "",
86
  "features": {
87
  "id": {
88
  "dtype": "string",
89
- "id": null,
90
  "_type": "Value"
91
  },
92
  "title": {
93
  "dtype": "string",
94
- "id": null,
95
  "_type": "Value"
96
  },
97
  "context": {
98
  "dtype": "string",
99
- "id": null,
100
  "_type": "Value"
101
  },
102
  "question": {
103
  "dtype": "string",
104
- "id": null,
105
  "_type": "Value"
106
  },
107
  "answers": {
108
  "feature": {
109
  "text": {
110
  "dtype": "string",
111
- "id": null,
112
  "_type": "Value"
113
  },
114
  "answer_start": {
115
  "dtype": "int32",
116
- "id": null,
117
  "_type": "Value"
118
  }
119
  },
120
- "length": -1,
121
- "id": null,
122
  "_type": "Sequence"
123
  },
124
  "metadata": {
125
  "split": {
126
  "dtype": "string",
127
- "id": null,
128
  "_type": "Value"
129
  },
130
  "model_in_the_loop": {
131
  "dtype": "string",
132
- "id": null,
133
  "_type": "Value"
134
  }
135
  }
136
  },
137
- "post_processed": null,
138
- "supervised_keys": null,
139
  "builder_name": "adversarial_qa",
 
140
  "config_name": "dbidaf",
141
  "version": {
142
  "version_str": "1.0.0",
143
- "description": null,
144
  "major": 1,
145
  "minor": 0,
146
  "patch": 0
@@ -148,33 +136,26 @@
148
  "splits": {
149
  "train": {
150
  "name": "train",
151
- "num_bytes": 9282518,
152
  "num_examples": 10000,
153
- "dataset_name": "adversarial_qa"
154
  },
155
  "validation": {
156
  "name": "validation",
157
- "num_bytes": 917943,
158
  "num_examples": 1000,
159
- "dataset_name": "adversarial_qa"
160
  },
161
  "test": {
162
  "name": "test",
163
- "num_bytes": 947111,
164
  "num_examples": 1000,
165
- "dataset_name": "adversarial_qa"
166
- }
167
- },
168
- "download_checksums": {
169
- "https://adversarialqa.github.io/data/aqa_v1.0.zip": {
170
- "num_bytes": 9018914,
171
- "checksum": "f4f3c23224a5060b28c35e35581bd5cf46256dda3665418fb83d036d0e0c93cf"
172
  }
173
  },
174
- "download_size": 9018914,
175
- "post_processing_size": null,
176
- "dataset_size": 11147572,
177
- "size_in_bytes": 20166486
178
  },
179
  "dbert": {
180
  "description": "AdversarialQA is a Reading Comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles using an adversarial model-in-the-loop.\nWe use three different models; BiDAF (Seo et al., 2016), BERT-Large (Devlin et al., 2018), and RoBERTa-Large (Liu et al., 2019) in the annotation loop and construct three datasets; D(BiDAF), D(BERT), and D(RoBERTa), each with 10,000 training examples, 1,000 validation, and 1,000 test examples.\nThe adversarial human annotation paradigm ensures that these datasets consist of questions that current state-of-the-art models (at least the ones used as adversaries in the annotation loop) find challenging.\n",
 
80
  },
81
  "dbidaf": {
82
  "description": "AdversarialQA is a Reading Comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles using an adversarial model-in-the-loop.\nWe use three different models; BiDAF (Seo et al., 2016), BERT-Large (Devlin et al., 2018), and RoBERTa-Large (Liu et al., 2019) in the annotation loop and construct three datasets; D(BiDAF), D(BERT), and D(RoBERTa), each with 10,000 training examples, 1,000 validation, and 1,000 test examples.\nThe adversarial human annotation paradigm ensures that these datasets consist of questions that current state-of-the-art models (at least the ones used as adversaries in the annotation loop) find challenging.\n",
83
+ "citation": "@article{bartolo2020beat,\n author = {Bartolo, Max and Roberts, Alastair and Welbl, Johannes and Riedel, Sebastian and Stenetorp, Pontus},\n title = {Beat the AI: Investigating Adversarial Human Annotation for Reading Comprehension},\n journal = {Transactions of the Association for Computational Linguistics},\n volume = {8},\n number = {},\n pages = {662-678},\n year = {2020},\n doi = {10.1162/tacl_a_00338},\n URL = { https://doi.org/10.1162/tacl_a_00338 },\n eprint = { https://doi.org/10.1162/tacl_a_00338 },\n abstract = { Innovations in annotation methodology have been a catalyst for Reading Comprehension (RC) datasets and models. One recent trend to challenge current RC models is to involve a model in the annotation process: Humans create questions adversarially, such that the model fails to answer them correctly. In this work we investigate this annotation methodology and apply it in three different settings, collecting a total of 36,000 samples with progressively stronger models in the annotation loop. This allows us to explore questions such as the reproducibility of the adversarial effect, transfer from data collected with varying model-in-the-loop strengths, and generalization to data collected without a model. We find that training on adversarially collected samples leads to strong generalization to non-adversarially collected datasets, yet with progressive performance deterioration with increasingly stronger models-in-the-loop. Furthermore, we find that stronger models can still learn from datasets collected with substantially weaker models-in-the-loop. When trained on data collected with a BiDAF model in the loop, RoBERTa achieves 39.9F1 on questions that it cannot answer when trained on SQuAD\u2014only marginally lower than when trained on data collected using RoBERTa itself (41.0F1). }\n}\n",
84
  "homepage": "https://adversarialqa.github.io/",
85
  "license": "",
86
  "features": {
87
  "id": {
88
  "dtype": "string",
 
89
  "_type": "Value"
90
  },
91
  "title": {
92
  "dtype": "string",
 
93
  "_type": "Value"
94
  },
95
  "context": {
96
  "dtype": "string",
 
97
  "_type": "Value"
98
  },
99
  "question": {
100
  "dtype": "string",
 
101
  "_type": "Value"
102
  },
103
  "answers": {
104
  "feature": {
105
  "text": {
106
  "dtype": "string",
 
107
  "_type": "Value"
108
  },
109
  "answer_start": {
110
  "dtype": "int32",
 
111
  "_type": "Value"
112
  }
113
  },
 
 
114
  "_type": "Sequence"
115
  },
116
  "metadata": {
117
  "split": {
118
  "dtype": "string",
 
119
  "_type": "Value"
120
  },
121
  "model_in_the_loop": {
122
  "dtype": "string",
 
123
  "_type": "Value"
124
  }
125
  }
126
  },
 
 
127
  "builder_name": "adversarial_qa",
128
+ "dataset_name": "adversarial_qa",
129
  "config_name": "dbidaf",
130
  "version": {
131
  "version_str": "1.0.0",
 
132
  "major": 1,
133
  "minor": 0,
134
  "patch": 0
 
136
  "splits": {
137
  "train": {
138
  "name": "train",
139
+ "num_bytes": 9282482,
140
  "num_examples": 10000,
141
+ "dataset_name": null
142
  },
143
  "validation": {
144
  "name": "validation",
145
+ "num_bytes": 917907,
146
  "num_examples": 1000,
147
+ "dataset_name": null
148
  },
149
  "test": {
150
  "name": "test",
151
+ "num_bytes": 946947,
152
  "num_examples": 1000,
153
+ "dataset_name": null
 
 
 
 
 
 
154
  }
155
  },
156
+ "download_size": 2721341,
157
+ "dataset_size": 11147336,
158
+ "size_in_bytes": 13868677
 
159
  },
160
  "dbert": {
161
  "description": "AdversarialQA is a Reading Comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles using an adversarial model-in-the-loop.\nWe use three different models; BiDAF (Seo et al., 2016), BERT-Large (Devlin et al., 2018), and RoBERTa-Large (Liu et al., 2019) in the annotation loop and construct three datasets; D(BiDAF), D(BERT), and D(RoBERTa), each with 10,000 training examples, 1,000 validation, and 1,000 test examples.\nThe adversarial human annotation paradigm ensures that these datasets consist of questions that current state-of-the-art models (at least the ones used as adversaries in the annotation loop) find challenging.\n",
dbidaf/test-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:18e15a1cd567d22adde6409d6507ecfc6c9aba04005d514ea9d58c95f9543b90
3
+ size 221193
dbidaf/train-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:164745ce3ef4098f71767eacf2cb3d120fd25b4f8593eb556a6f5db7224fcfbd
3
+ size 2266212
dbidaf/validation-00000-of-00001.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3ee96cb95025f2c94c450b20de2b3540593823252f40d54a7cb1b33d24e75002
3
+ size 233936