MicPie commited on
Commit
55c1c98
1 Parent(s): 7046d3a

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +81 -39
README.md CHANGED
@@ -38,7 +38,6 @@ task_ids:
38
  - tabular-multi-label-classification
39
  ---
40
 
41
-
42
  # Dataset Card for "AdapTable-full" - Dataset of Few-shot Tasks from Tables
43
 
44
  ## Table of Contents
@@ -74,39 +73,81 @@ task_ids:
74
 
75
  ### Dataset Summary
76
 
77
- The AdapTable dataset consists of tables that naturally occur on the web, that are formatted as few-shot tasks for fine-tuning language models to improve their few-shot performance.
78
-
79
- Starting from the initial WTC corpus of 50M tables, we apply our tables-to-tasks procedure to produce our resulting dataset, [AdapTable-full](https://huggingface.co/datasets/MicPie/adaptable_full), which comprises 413,350 tasks from 23,744 unique websites.
80
-
81
- We additionally export a version [AdapTable-unique](https://huggingface.co/datasets/MicPie/adaptable_unique) for comparison. This is the same as [AdapTable-full](https://huggingface.co/datasets/MicPie/adaptable_full) but filtered to have a maximum of one task per website. AdapTable-unique contains exactly 23,744 tasks from 23,744 websites.
82
-
83
- Furthermore, there is the [AdapTable-diverse](https://huggingface.co/datasets/MicPie/adaptable_diverse) dataset available as well as different AdapTable data subsets based on the website origin:
84
- * [AdapTable-baseball.fantasysports.yahoo.com](https://huggingface.co/datasets/MicPie/adaptable_baseball.fantasysports.yahoo.com)
85
- * [AdapTable-bulbapedia.bulbagarden.net](https://huggingface.co/datasets/MicPie/adaptable_bulbapedia.bulbagarden.net)
86
- * [AdapTable-cappex.com](https://huggingface.co/datasets/MicPie/adaptable_cappex.com)
87
- * [AdapTable-cram.com](https://huggingface.co/datasets/MicPie/adaptable_cram.com)
88
- * [AdapTable-dividend.com](https://huggingface.co/datasets/MicPie/adaptable_dividend.com)
89
- * [AdapTable-dummies.com](https://huggingface.co/datasets/MicPie/adaptable_dummies.com)
90
- * [AdapTable-en.wikipedia.org](https://huggingface.co/datasets/MicPie/adaptable_en.wikipedia.org)
91
- * [AdapTable-ensembl.org](https://huggingface.co/datasets/MicPie/adaptable_ensembl.org)
92
- * [AdapTable-gamefaqs.com](https://huggingface.co/datasets/MicPie/adaptable_gamefaqs.com)
93
- * [AdapTable-mgoblog.com](https://huggingface.co/datasets/MicPie/adaptable_mgoblog.com)
94
- * [AdapTable-mmo-champion.com](https://huggingface.co/datasets/MicPie/adaptable_mmo-champion.com)
95
- * [AdapTable-msdn.microsoft.com](https://huggingface.co/datasets/MicPie/adaptable_msdn.microsoft.com)
96
- * [AdapTable-phonearena.com](https://huggingface.co/datasets/MicPie/adaptable_phonearena.com)
97
- * [AdapTable-sittercity.com](https://huggingface.co/datasets/MicPie/adaptable_sittercity.com)
98
- * [AdapTable-sporcle.com](https://huggingface.co/datasets/MicPie/adaptable_sporcle.com)
99
- * [AdapTable-studystack.com](https://huggingface.co/datasets/MicPie/adaptable_studystack.com)
100
- * [AdapTable-support.google.com](https://huggingface.co/datasets/MicPie/adaptable_support.google.com)
101
- * [AdapTable-w3.org](https://huggingface.co/datasets/MicPie/adaptable_w3.org)
102
- * [AdapTable-wiki.openmoko.org](https://huggingface.co/datasets/MicPie/adaptable_wiki.openmoko.org)
103
- * [AdapTable-wkdu.org](https://huggingface.co/datasets/MicPie/adaptable_wkdu.org)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
104
 
105
  ### Supported Tasks and Leaderboards
106
 
107
- Since the tables come from the web, the distribution of tasks and topics is very broad. The shape of our dataset is very wide, i.e., we have 1000's tasks, while each task has only a few examples, compared to most current NLP datasets which are very deep, i.e., 10s of tasks with many examples. This implies that our dataset covers a broad range of potential tasks, e.g., multiple-choice, question-answering, table-question-answering, text-classification, etc.
108
 
109
- The intended use of this dataset is to improve few-shot performance by finetuning/pretraining onour dataset.
110
 
111
  ### Languages
112
 
@@ -124,11 +165,11 @@ There are also additional meta-data fields such as 'pageTitle', 'title', 'output
124
 
125
  'task': task identifier
126
 
127
- 'input': column elements of a specific row in table.
128
 
129
  'options': for multiple choice classification, it provides the options to choose from.
130
 
131
- 'output': target column element of same row as input.
132
 
133
  'pageTitle': the title of the page containing the table.
134
 
@@ -140,7 +181,7 @@ There are also additional meta-data fields such as 'pageTitle', 'title', 'output
140
 
141
  ### Data Splits
142
 
143
- AdapTable-full does not come with additional data splits.
144
 
145
  ## Dataset Creation
146
 
@@ -149,13 +190,13 @@ AdapTable-full does not come with additional data splits.
149
  How do we convert tables to few-shot tasks?
150
  Unlike unstructured text, structured data in the form of tables lends itself easily to the few-shot task format. Given a table where each row is an instance of a similar class and the columns describe the attributes of each instance, we can turn each row into a task example to predict one attribute given the others. When the table has more than one row, we instantly have multiple examples of this task by using each row as a single example, and thus each table becomes a few-shot dataset for a particular task.
151
 
152
- The few-shot setting in this setting is significant: Tables often do not come with clear instructions for each field, so tasks may be underspecified if prompted in a zero-shot manner, but the intended task becomes clearer when examples are provided. This makes a good two-way match: The few-shot format is a perfect setup for table learning, and tables provide a natural dataset for few-shot training.
153
 
154
  ### Source Data
155
 
156
  #### Initial Data Collection and Normalization
157
 
158
- The data processing pipelines is explained in detail in section 2.3 of our publication
159
 
160
  #### Who are the source language producers?
161
 
@@ -166,10 +207,11 @@ The dataset is extracted from [WDC Web Table Corpora](http://webdatacommons.org/
166
  #### Annotation process
167
 
168
  No manual annotation process used.
 
169
 
170
  #### Who are the annotators?
171
 
172
- n/a
173
 
174
  ### Personal and Sensitive Information
175
 
@@ -181,12 +223,12 @@ The data was extracted from [WDC Web Table Corpora](http://webdatacommons.org/we
181
 
182
  The purpose of this dataset is to help develop models that are better at few-shot learning and have higher few-shot performance by fine-tuning few-shot tasks extracted from tables.
183
 
184
- While tables have a similar structure to few-shot tasks and we do see an improved performance on few-shot tasks in our paper, we want to make clear that finetuning on tables also has its risks. First of all, since the tables are extracted from the web, they may contain user identities or otherwise sensitive information which a model might reveal at inference, or which could influence the learning process of a model in a negative way. Second, since tables are very diverse in nature, the model also trains on low-quality data or data with an unusual structure. While it is interesting that training on such data improves few-shot performance on downstream tasks, this could also imply that the model learns concepts that are very dissimilar to human concepts that would be useful for a certain downstream task. In other words, it is possible that the model learns weird things that are helpful on the evaluated downstream tasks, but might lead to bad out-of-distribution behavior.
185
 
186
  ### Discussion of Biases
187
 
188
- Since our dataset contains tables that are scraped from the web, it will also contain many toxic, racist, sexist, and otherwise harmful biases and texts. We have not run any analysis on the biases prevalent in our datasets. Neither have we explicitly filtered the content for toxic content.
189
- This implies that a model trained on our dataset will reinforce harmful biases and toxic text that exist in our dataset.
190
 
191
  ### Other Known Limitations
192
 
 
38
  - tabular-multi-label-classification
39
  ---
40
 
 
41
  # Dataset Card for "AdapTable-full" - Dataset of Few-shot Tasks from Tables
42
 
43
  ## Table of Contents
 
73
 
74
  ### Dataset Summary
75
 
76
+ The AdapTable dataset consists of tables that naturally occur on the web and that are formatted as few-shot tasks for fine-tuning language models to improve their few-shot performance.
77
+
78
+ There are several dataset versions available:
79
+
80
+ * [AdapTable-full](https://huggingface.co/datasets/MicPie/adaptable_full): Starting from the initial WTC corpus of 50M tables, we apply our tables-to-tasks procedure to produce our resulting dataset, [AdapTable-full](https://huggingface.co/datasets/MicPie/adaptable_full), which comprises 413,350 tasks from 23,744 unique websites.
81
+
82
+ * [AdapTable-unique](https://huggingface.co/datasets/MicPie/adaptable_unique): This is the same as [AdapTable-full](https://huggingface.co/datasets/MicPie/adaptable_full) but filtered to have a maximum of one task per website. [AdapTable-unique](https://huggingface.co/datasets/MicPie/adaptable_unique) contains exactly 23,744 tasks from 23,744 websites.
83
+
84
+ * [AdapTable-5k](https://huggingface.co/datasets/MicPie/adaptable_5k): This dataset uses 5k random tables from the full dataset.
85
+
86
+ * AdapTable data subsets based on a manual human quality rating:
87
+ * [AdapTable-rated-low](https://huggingface.co/datasets/MicPie/adaptable_rated-low)
88
+ * [AdapTable-rated-medium](https://huggingface.co/datasets/MicPie/adaptable_rated-medium)
89
+ * [AdapTable-rated-high](https://huggingface.co/datasets/MicPie/adaptable_rated-high)
90
+
91
+ * AdapTable data subsets based on the website of origin:
92
+ * [AdapTable-baseball.fantasysports.yahoo.com](https://huggingface.co/datasets/MicPie/adaptable_baseball.fantasysports.yahoo.com)
93
+ * [AdapTable-bulbapedia.bulbagarden.net](https://huggingface.co/datasets/MicPie/adaptable_bulbapedia.bulbagarden.net)
94
+ * [AdapTable-cappex.com](https://huggingface.co/datasets/MicPie/adaptable_cappex.com)
95
+ * [AdapTable-cram.com](https://huggingface.co/datasets/MicPie/adaptable_cram.com)
96
+ * [AdapTable-dividend.com](https://huggingface.co/datasets/MicPie/adaptable_dividend.com)
97
+ * [AdapTable-dummies.com](https://huggingface.co/datasets/MicPie/adaptable_dummies.com)
98
+ * [AdapTable-en.wikipedia.org](https://huggingface.co/datasets/MicPie/adaptable_en.wikipedia.org)
99
+ * [AdapTable-ensembl.org](https://huggingface.co/datasets/MicPie/adaptable_ensembl.org)
100
+ * [AdapTable-gamefaqs.com](https://huggingface.co/datasets/MicPie/adaptable_gamefaqs.com)
101
+ * [AdapTable-mgoblog.com](https://huggingface.co/datasets/MicPie/adaptable_mgoblog.com)
102
+ * [AdapTable-mmo-champion.com](https://huggingface.co/datasets/MicPie/adaptable_mmo-champion.com)
103
+ * [AdapTable-msdn.microsoft.com](https://huggingface.co/datasets/MicPie/adaptable_msdn.microsoft.com)
104
+ * [AdapTable-phonearena.com](https://huggingface.co/datasets/MicPie/adaptable_phonearena.com)
105
+ * [AdapTable-sittercity.com](https://huggingface.co/datasets/MicPie/adaptable_sittercity.com)
106
+ * [AdapTable-sporcle.com](https://huggingface.co/datasets/MicPie/adaptable_sporcle.com)
107
+ * [AdapTable-studystack.com](https://huggingface.co/datasets/MicPie/adaptable_studystack.com)
108
+ * [AdapTable-support.google.com](https://huggingface.co/datasets/MicPie/adaptable_support.google.com)
109
+ * [AdapTable-w3.org](https://huggingface.co/datasets/MicPie/adaptable_w3.org)
110
+ * [AdapTable-wiki.openmoko.org](https://huggingface.co/datasets/MicPie/adaptable_wiki.openmoko.org)
111
+ * [AdapTable-wkdu.org](https://huggingface.co/datasets/MicPie/adaptable_wkdu.org)
112
+
113
+ * AdapTable data subsets based on clustering (for the clustering details please see our publication):
114
+ * [AdapTable-cluster00](https://huggingface.co/datasets/MicPie/adaptable_cluster00)
115
+ * [AdapTable-cluster01](https://huggingface.co/datasets/MicPie/adaptable_cluster01)
116
+ * [AdapTable-cluster02](https://huggingface.co/datasets/MicPie/adaptable_cluster02)
117
+ * [AdapTable-cluster03](https://huggingface.co/datasets/MicPie/adaptable_cluster03)
118
+ * [AdapTable-cluster04](https://huggingface.co/datasets/MicPie/adaptable_cluster04)
119
+ * [AdapTable-cluster05](https://huggingface.co/datasets/MicPie/adaptable_cluster05)
120
+ * [AdapTable-cluster06](https://huggingface.co/datasets/MicPie/adaptable_cluster06)
121
+ * [AdapTable-cluster07](https://huggingface.co/datasets/MicPie/adaptable_cluster07)
122
+ * [AdapTable-cluster08](https://huggingface.co/datasets/MicPie/adaptable_cluster08)
123
+ * [AdapTable-cluster09](https://huggingface.co/datasets/MicPie/adaptable_cluster09)
124
+ * [AdapTable-cluster10](https://huggingface.co/datasets/MicPie/adaptable_cluster10)
125
+ * [AdapTable-cluster11](https://huggingface.co/datasets/MicPie/adaptable_cluster11)
126
+ * [AdapTable-cluster12](https://huggingface.co/datasets/MicPie/adaptable_cluster12)
127
+ * [AdapTable-cluster13](https://huggingface.co/datasets/MicPie/adaptable_cluster13)
128
+ * [AdapTable-cluster14](https://huggingface.co/datasets/MicPie/adaptable_cluster14)
129
+ * [AdapTable-cluster15](https://huggingface.co/datasets/MicPie/adaptable_cluster15)
130
+ * [AdapTable-cluster16](https://huggingface.co/datasets/MicPie/adaptable_cluster16)
131
+ * [AdapTable-cluster17](https://huggingface.co/datasets/MicPie/adaptable_cluster17)
132
+ * [AdapTable-cluster18](https://huggingface.co/datasets/MicPie/adaptable_cluster18)
133
+ * [AdapTable-cluster19](https://huggingface.co/datasets/MicPie/adaptable_cluster19)
134
+ * [AdapTable-cluster20](https://huggingface.co/datasets/MicPie/adaptable_cluster20)
135
+ * [AdapTable-cluster21](https://huggingface.co/datasets/MicPie/adaptable_cluster21)
136
+ * [AdapTable-cluster22](https://huggingface.co/datasets/MicPie/adaptable_cluster22)
137
+ * [AdapTable-cluster23](https://huggingface.co/datasets/MicPie/adaptable_cluster23)
138
+ * [AdapTable-cluster24](https://huggingface.co/datasets/MicPie/adaptable_cluster24)
139
+ * [AdapTable-cluster25](https://huggingface.co/datasets/MicPie/adaptable_cluster25)
140
+ * [AdapTable-cluster26](https://huggingface.co/datasets/MicPie/adaptable_cluster26)
141
+ * [AdapTable-cluster27](https://huggingface.co/datasets/MicPie/adaptable_cluster27)
142
+ * [AdapTable-cluster28](https://huggingface.co/datasets/MicPie/adaptable_cluster28)
143
+ * [AdapTable-cluster29](https://huggingface.co/datasets/MicPie/adaptable_cluster29)
144
+ * [AdapTable-cluster-noise](https://huggingface.co/datasets/MicPie/adaptable_cluster-noise)
145
 
146
  ### Supported Tasks and Leaderboards
147
 
148
+ Since the tables come from the web, the distribution of tasks and topics is very broad. The shape of our dataset is very wide, i.e., we have 1000's of tasks, while each task has only a few examples, compared to most current NLP datasets which are very deep, i.e., 10s of tasks with many examples. This implies that our dataset covers a broad range of potential tasks, e.g., multiple-choice, question-answering, table-question-answering, text-classification, etc.
149
 
150
+ The intended use of this dataset is to improve few-shot performance by fine-tuning/pre-training on our dataset.
151
 
152
  ### Languages
153
 
 
165
 
166
  'task': task identifier
167
 
168
+ 'input': column elements of a specific row in the table.
169
 
170
  'options': for multiple choice classification, it provides the options to choose from.
171
 
172
+ 'output': target column element of the same row as input.
173
 
174
  'pageTitle': the title of the page containing the table.
175
 
 
181
 
182
  ### Data Splits
183
 
184
+ The AdapTable datasets do not come with additional data splits.
185
 
186
  ## Dataset Creation
187
 
 
190
  How do we convert tables to few-shot tasks?
191
  Unlike unstructured text, structured data in the form of tables lends itself easily to the few-shot task format. Given a table where each row is an instance of a similar class and the columns describe the attributes of each instance, we can turn each row into a task example to predict one attribute given the others. When the table has more than one row, we instantly have multiple examples of this task by using each row as a single example, and thus each table becomes a few-shot dataset for a particular task.
192
 
193
+ The few-shot setting in this setup is significant: Tables often do not come with clear instructions for each field, so tasks may be underspecified if prompted in a zero-shot manner, but the intended task becomes clearer when examples are provided. This makes a good two-way match: The few-shot format is a perfect setup for table learning, and tables provide a natural dataset for few-shot training.
194
 
195
  ### Source Data
196
 
197
  #### Initial Data Collection and Normalization
198
 
199
+ The data processing pipeline is explained in detail in our publication.
200
 
201
  #### Who are the source language producers?
202
 
 
207
  #### Annotation process
208
 
209
  No manual annotation process used.
210
+ Only for the [AdapTable-rated-low](https://huggingface.co/datasets/MicPie/adaptable_rated-low), [AdapTable-rated-medium](https://huggingface.co/datasets/MicPie/adaptable_rated-medium), and [AdapTable-rated-high](https://huggingface.co/datasets/MicPie/adaptable_rated-high) manual annotations were carried out.
211
 
212
  #### Who are the annotators?
213
 
214
+ People involved in the publication.
215
 
216
  ### Personal and Sensitive Information
217
 
 
223
 
224
  The purpose of this dataset is to help develop models that are better at few-shot learning and have higher few-shot performance by fine-tuning few-shot tasks extracted from tables.
225
 
226
+ While tables have a similar structure to few-shot tasks and we do see an improved performance on few-shot tasks in our paper, we want to make clear that fine-tuning on tables also has its risks. First of all, since the tables are extracted from the web, they may contain user identities or otherwise sensitive information which a model might reveal at inference, or which could influence the learning process of a model in a negative way. Second, since tables are very diverse in nature, the model also trains on low-quality data or data with an unusual structure. While it is interesting that training on such data improves few-shot performance on downstream tasks, this could also imply that the model learns concepts that are very dissimilar to human concepts that would be useful for a certain downstream task. In other words, it is possible that the model learns weird things that are helpful on the evaluated downstream tasks, but might lead to bad out-of-distribution behavior.
227
 
228
  ### Discussion of Biases
229
 
230
+ Since our dataset contains tables that are scraped from the web, it will also contain many toxic, racist, sexist, and otherwise harmful biases and texts. We have not run any analysis on the biases prevalent in our datasets. Neither have we explicitly filtered the content.
231
+ This implies that a model trained on our dataset will potentially reinforce harmful biases and toxic text that exist in our dataset.
232
 
233
  ### Other Known Limitations
234