Datasets:

Size Categories:
n>1T
ArXiv:
License:
drschwenk commited on
Commit
e2d81d9
1 Parent(s): 85601d1

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +40 -47
README.md CHANGED
@@ -12,11 +12,12 @@ size_categories:
12
 
13
  [MADLAD-400 (*Multilingual Audited Dataset: Low-resource And Document-level*)](https://arxiv.org/abs/2309.04662) is
14
  a document-level multilingual dataset based on Common Crawl, covering 419
15
- languages in total. We use all snapshots of CommonCrawl available as of August
16
  1, 2022. The primary advantage of this dataset over similar datasets is that it
17
  is more multilingual (419 languages), it is audited and more highly filtered,
18
  and it is document-level. The main disadvantage is also its strength -- being
19
- more filtered, it may lack the recall needed for some applications.
 
20
 
21
  There are two versions released: the **noisy** dataset, which has no filtering
22
  except document-level LangID, and the **clean** dataset, which has a variety of
@@ -57,8 +58,8 @@ the hopes that it will increase robustness to web-domain text.
57
 
58
  ## Filtering
59
 
60
- Before separating the raw CommonCrawl corpus by LangID, we carry out the
61
- following filtering steps as done by Raffel et al (2020):
62
 
63
  - Discarded any page with fewer than 5 sentences and only retained lines that
64
  contained at least 3 words.
@@ -66,8 +67,8 @@ following filtering steps as done by Raffel et al (2020):
66
  - Removed any page where the phrase “lorem ipsum” appeared.
67
  - Removed any pages containing the phrases "terms of use", "privacy policy",
68
  "cookie policy", "uses cookies", "use of cookies", "use cookies"
69
- - We removed any pages that contained a curly bracket.
70
- - To deduplicate the data set, we discarded all but one of any three-sentence span occurring more than once in the data set.
71
 
72
  The `noisy` subset of the data was filtered only by document-level LangID, which
73
  was taken to be the majority sentence-level LangID prediction. The `clean`
@@ -90,7 +91,7 @@ of the following were true:
90
 
91
  ### Cursed Substrings
92
 
93
- Based on the initial round of data audits, we collected a heuristic list of
94
  substrings and regexes accounting for a large amount of questionable content.
95
  Keep in mind that these all are fed into the `pct_questionable` score -- a
96
  sentence is only excluded from the `clean` dataset if over 20% of the sentences
@@ -118,12 +119,12 @@ CURSED_SUBSTRINGS = [" №", "���", "\\|\\s*$", " nr\\.$", "aute irure dol
118
 
119
  Many languages using Brahmic Abugida (South and Southeast Asian scripts like
120
  Devanagari, Khmer, etc.) use some variant on the virama character. For whatever
121
- reason, we found that this character was often messed up in the common crawl
122
- data we were Crawling. Therefore, for the languages `bn my pa gu or ta te kn ml
123
  si th tl mn lo bo km hi mr ne gom as jv dv bho dz hne ks_Deva mag mni shn yue zh
124
- ja kjg mnw ksw rki mtr mwr xnr`, we did a special correction step.
125
 
126
- For these languages, we took the list of all virama characters and removed all
127
  unnecessary spaces between each instance of a virama character and the next
128
  character with a regex.
129
 
@@ -134,7 +135,7 @@ character with a regex.
134
  ### Myanmar Font Compatibility
135
 
136
  Prior to 2019, the most popular font for Burmese websites was the Zawgyi font.
137
- We used [Myanmar Tools](https://github.com/google/myanmar-tools) to convert text
138
 
139
  Several scripts, like the Chinese script, Tibetan script, and Thai, do not use
140
  whitespace to separate characters. The languages with this property in this
@@ -148,67 +149,58 @@ see below.)
148
 
149
  ### Special filters
150
 
151
- Chinese had a particular issue with pornographic content. We combed through the
152
- data and developed a list of strings likely to be present in pornographic
153
- content, and filtered out all pages containing at least one of these strings.
154
- Resulted in 17% reduction in number of documents and 56% reduction in file size.
155
 
156
  ```
157
  pornsignals = "caoporn caoprom caopron caoporen caoponrn caoponav caopom caoorn 99re dy888 caopro hezyo re99 4438x zooskool xfplay 7tav xxoo xoxo 52av freexx 91chinese anquye cao97 538porm 87fuli 91pron 91porn 26uuu 4438x 182tv kk4444 777me ae86 91av 720lu yy6080 6080yy qqchub paa97 aiai777 yy4480 videossexo 91free 一级特黄大片 偷拍久久国产视频 日本毛片免费视频观看 久久免费热在线精品 高清毛片在线看 日本毛片高清免费视频 一级黄色录像影片 亚洲男人天堂 久久精品视频在线看 自拍区偷拍亚洲视频 亚洲人成视频在线播放 色姑娘综合站 丁香五月啪啪 在线视频成人社区 亚洲人成视频在线播放 久久国产自偷拍 一本道 大香蕉无码 香港经典三级 亚洲成在人线免费视频 天天色综合网 大香蕉伊人久草 欧美一级高清片 天天鲁夜夜啪视频在线 免费黄片视频在线观看 加比勒久久综合 久草热久草在线视频 韩国三级片大全在线观看 青青草在线视频 美国一级毛片 久草在线福利资源 啪啪啪视频在线观看免费 成人福利视频在线观看 婷婷我去也 老司机在线国产 久久成人视频 手机看片福利永久国产 高清国产偷拍在线 大香蕉在线影院 日本高清免费一本视频 男人的天堂东京热 影音先锋男人资源 五月婷婷开心中文字幕 亚洲香蕉视频在线播放 天天啪久久爱视频精品 超碰久久人人摸人人搞".split()
158
  ```
159
 
160
- ## Language code notes
161
-
162
- All these different datasets have slightly different language codes! We use the
163
- BCP-47 standard, which specifies the 2-letter ISO-693-1 code when applicable,
164
- and otherwise the ISO-693-3 code. Script tags and region tags are omitted when
165
- they are defined as the default value by CLDR, and otherwise included; for
166
- instance `ks` refers to Kashmiri in Nastaliq/Arabic script (CLDR default),
167
- whereas `ks_Deva` refers to Kashmiri in Devanagari.
168
-
169
  A few more random notes, comparing to common alternative codes for these
170
  languages:
171
 
172
- * We use `fil` for Filipino/Tagalog, not `tl`
173
- * We use `ak` for Twi/Akan, rather than `tw`. This includes Fante.
174
- * Unfortunately, we use the macro code `chm` for Meadow Mari (instead of the
175
  correct `mhr`), and `mrj` for Hill Mari
176
- * By convention, we use `no` for Norwegian Bokmål, whereas some resources use
177
  `nb`
178
- * By convention we use `ps` for Pashto instead of `pbt` (Southern Pashto)
179
- * By convention, we use `ms` for Standard Malay, not `zlm`
180
- * By convention, we use `sq` for Albanian, and don't distinguish dialects like
181
  Gheg (`aln`) and Tosk (`als`)
182
- * We use `ber` as the code for Tamazight, after consultation with Tamazight
183
  speakers opining that the dialect distinctions are not significant. Other
184
  resources use the individual codes like `tzm` and `kab`.
185
- * We use the macrocode `qu` for Quechua. In practice, this seems usually to be
186
  a mix of the Ayacucho and Cusco dialects. Other resources, like NLLB, may
187
  use the dialect code, e.g. `quy` for Ayacucho Chanka. The same is true for a
188
  few other macro codes, like `ff` (Macro code for Fulfulde, whereas other
189
  sources may use e.g. `fuv`.)
190
  * Really, there are notes that can be made about almost any code, from the
191
  well-accepted conventions like `zh` for Mandarin, to many dialectical notes,
192
- like which variant of Hmong really is the `hmn` data? But The above ones are
193
- made specifically for ones where we are aware of other datasources floating
194
  out there that use different conventions.
195
 
 
196
  ## Audit
197
 
198
- Following [Quality at a Glance](https://arxiv.org/abs/2103.12028), we performed
199
- an "audit" of every corpus in this dataset. Although we did not speak most
200
- languages, we were able to give high-level comments on the general quality. We
201
  looked at a sample of 20 documents of each language.
202
 
203
- After an initial round of auditing, we devised a new set of filters and applied
204
- them. We then re-did all audits.
205
 
206
  ### Overall notes from the audit
207
 
208
- Our decision was to **include languages that looked noisy, but omit any language
209
  that was clearly majority noise, or only had 20 or fewer docs.** This is a low
210
- bar -- twenty documents can be very little indeed, and some of the corpora we
211
- release are quite noisy, but all of them should have at least the potential to
212
  be used in some useful way. The motivation for not releasing nonsense or tiny
213
  datasets is to not give a false sense of how multilingual this dataset actually
214
  is ("Representation washing"), as recommended by **Quality at a Glance**.
@@ -222,6 +214,7 @@ A few overarching points:
222
  * Indian languages in the Latin script had a high concentration of
223
  pornographic content.
224
 
 
225
  ### Renames and Merges as a result of the Audit
226
 
227
  In several cases, it was clear from the audit that the corpora were not in the
@@ -238,7 +231,7 @@ renames:
238
  * `bjj` merged into the `awa` dataset
239
 
240
  ## Canaries
241
- We provide canaries in separate `canaries` folder. Canaries are organized into three directions: `monolingual` hosts canaries designed for the MADLAD-400 monody data, `multiway` for the multiway data, and `generic` the generic canaries generated only from the model's vocabulary.
242
 
243
  * Monolingual: Canaries here are organized by the language the canary was generated from. This corresponds exactly to the `translate_copy` setting in the paper, where the source and target language match.
244
 
@@ -247,7 +240,7 @@ We provide canaries in separate `canaries` folder. Canaries are organized into t
247
  Within each subdirectory above, canaries are into separate files named by the canary type. There is always only a single file for each canary type. The `generic` folder contains within it the four canary types.
248
 
249
 
250
- Canaries can be mixed in with normal training data to then be analyzed post-hoc to training.
251
 
252
 
253
  ## References
@@ -265,7 +258,7 @@ This data is released with the `CC-BY-4.0` license.
265
 
266
  ## Detailed notes from the audit
267
 
268
- Here are the notes we gave on all languages, along with the number of documents
269
  found, and the final decision made with respect to including the language in
270
  this dataset.
271
 
 
12
 
13
  [MADLAD-400 (*Multilingual Audited Dataset: Low-resource And Document-level*)](https://arxiv.org/abs/2309.04662) is
14
  a document-level multilingual dataset based on Common Crawl, covering 419
15
+ languages in total. This uses all snapshots of CommonCrawl available as of August
16
  1, 2022. The primary advantage of this dataset over similar datasets is that it
17
  is more multilingual (419 languages), it is audited and more highly filtered,
18
  and it is document-level. The main disadvantage is also its strength -- being
19
+ more filtered, it may lack the recall needed for some applications.
20
+
21
 
22
  There are two versions released: the **noisy** dataset, which has no filtering
23
  except document-level LangID, and the **clean** dataset, which has a variety of
 
58
 
59
  ## Filtering
60
 
61
+ Before separating the raw CommonCrawl corpus by LangID, these
62
+ filtering steps are done, similar to Raffel et al (2020):
63
 
64
  - Discarded any page with fewer than 5 sentences and only retained lines that
65
  contained at least 3 words.
 
67
  - Removed any page where the phrase “lorem ipsum” appeared.
68
  - Removed any pages containing the phrases "terms of use", "privacy policy",
69
  "cookie policy", "uses cookies", "use of cookies", "use cookies"
70
+ - Removed any pages that contained a curly bracket.
71
+ - To deduplicate the data set, discarded all but one of any three-sentence span occurring more than once in the data set.
72
 
73
  The `noisy` subset of the data was filtered only by document-level LangID, which
74
  was taken to be the majority sentence-level LangID prediction. The `clean`
 
91
 
92
  ### Cursed Substrings
93
 
94
+ Based on the initial round of data audits, the authors created a heuristic list of
95
  substrings and regexes accounting for a large amount of questionable content.
96
  Keep in mind that these all are fed into the `pct_questionable` score -- a
97
  sentence is only excluded from the `clean` dataset if over 20% of the sentences
 
119
 
120
  Many languages using Brahmic Abugida (South and Southeast Asian scripts like
121
  Devanagari, Khmer, etc.) use some variant on the virama character. For whatever
122
+ reason, it was found that this character was often messed up in the common crawl
123
+ snapshots used. Therefore, for the languages `bn my pa gu or ta te kn ml
124
  si th tl mn lo bo km hi mr ne gom as jv dv bho dz hne ks_Deva mag mni shn yue zh
125
+ ja kjg mnw ksw rki mtr mwr xnr`, a special correction step was done.
126
 
127
+ For these languages, the authors took the list of all virama characters and removed all
128
  unnecessary spaces between each instance of a virama character and the next
129
  character with a regex.
130
 
 
135
  ### Myanmar Font Compatibility
136
 
137
  Prior to 2019, the most popular font for Burmese websites was the Zawgyi font.
138
+ The authors used [Myanmar Tools](https://github.com/google/myanmar-tools) to convert text.
139
 
140
  Several scripts, like the Chinese script, Tibetan script, and Thai, do not use
141
  whitespace to separate characters. The languages with this property in this
 
149
 
150
  ### Special filters
151
 
152
+ Chinese had a particular issue with pornographic content. After manual inspection
153
+ a list of strings likely to be present in pornographic content was developed. All
154
+ pages containing at least one of these strings were removed. Resulted in 17%
155
+ reduction in number of documents and 56% reduction in file size.
156
 
157
  ```
158
  pornsignals = "caoporn caoprom caopron caoporen caoponrn caoponav caopom caoorn 99re dy888 caopro hezyo re99 4438x zooskool xfplay 7tav xxoo xoxo 52av freexx 91chinese anquye cao97 538porm 87fuli 91pron 91porn 26uuu 4438x 182tv kk4444 777me ae86 91av 720lu yy6080 6080yy qqchub paa97 aiai777 yy4480 videossexo 91free 一级特黄大片 偷拍久久国产视频 日本毛片免费视频观看 久久免费热在线精品 高清毛片在线看 日本毛片高清免费视频 一级黄色录像影片 亚洲男人天堂 久久精品视频在线看 自拍区偷拍亚洲视频 亚洲人成视频在线播放 色姑娘综合站 丁香五月啪啪 在线视频成人社区 亚洲人成视频在线播放 久久国产自偷拍 一本道 大香蕉无码 香港经典三级 亚洲成在人线免费视频 天天色综合网 大香蕉伊人久草 欧美一级高清片 天天鲁夜夜啪视频在线 免费黄片视频在线观看 加比勒久久综合 久草热久草在线视频 韩国三级片大全在线观看 青青草在线视频 美国一级毛片 久草在线福利资源 啪啪啪视频在线观看免费 成人福利视频在线观看 婷婷我去也 老司机在线国产 久久成人视频 手机看片福利永久国产 高清国产偷拍在线 大香蕉在线影院 日本高清免费一本视频 男人的天堂东京热 影音先锋男人资源 五月婷婷开心中文字幕 亚洲香蕉视频在线播放 天天啪久久爱视频精品 超碰久久人人摸人人搞".split()
159
  ```
160
 
 
 
 
 
 
 
 
 
 
161
  A few more random notes, comparing to common alternative codes for these
162
  languages:
163
 
164
+ * `fil` for Filipino/Tagalog, not `tl`
165
+ * `ak` for Twi/Akan, rather than `tw`. This includes Fante.
166
+ * Unfortunately use the macro code `chm` for Meadow Mari (instead of the
167
  correct `mhr`), and `mrj` for Hill Mari
168
+ * `no` for Norwegian Bokmål, whereas some resources use
169
  `nb`
170
+ * `ps` for Pashto instead of `pbt` (Southern Pashto)
171
+ * `ms` for Standard Malay, not `zlm`
172
+ * `sq` for Albanian, and don't distinguish dialects like
173
  Gheg (`aln`) and Tosk (`als`)
174
+ * `ber` as the code for Tamazight, after consultation with Tamazight
175
  speakers opining that the dialect distinctions are not significant. Other
176
  resources use the individual codes like `tzm` and `kab`.
177
+ * Macrocode `qu` for Quechua. In practice, this seems usually to be
178
  a mix of the Ayacucho and Cusco dialects. Other resources, like NLLB, may
179
  use the dialect code, e.g. `quy` for Ayacucho Chanka. The same is true for a
180
  few other macro codes, like `ff` (Macro code for Fulfulde, whereas other
181
  sources may use e.g. `fuv`.)
182
  * Really, there are notes that can be made about almost any code, from the
183
  well-accepted conventions like `zh` for Mandarin, to many dialectical notes,
184
+ like which variant of Hmong really is the `hmn` data? But the above ones are
185
+ made specifically for ones where the authors are aware of other datasources floating
186
  out there that use different conventions.
187
 
188
+
189
  ## Audit
190
 
191
+ Following [Quality at a Glance](https://arxiv.org/abs/2103.12028), the authors performed
192
+ an "audit" of every corpus in this dataset. Although the authors did not speak most
193
+ languages, they were able to give high-level comments on the general quality. They
194
  looked at a sample of 20 documents of each language.
195
 
196
+ After an initial round of auditing, they devised a new set of filters and applied
197
+ them. They then re-did all audits.
198
 
199
  ### Overall notes from the audit
200
 
201
+ The decision was to **include languages that looked noisy, but omit any language
202
  that was clearly majority noise, or only had 20 or fewer docs.** This is a low
203
+ bar -- twenty documents can be very little indeed, and some of the corpora released are quite noisy, but all of them should have at least the potential to
 
204
  be used in some useful way. The motivation for not releasing nonsense or tiny
205
  datasets is to not give a false sense of how multilingual this dataset actually
206
  is ("Representation washing"), as recommended by **Quality at a Glance**.
 
214
  * Indian languages in the Latin script had a high concentration of
215
  pornographic content.
216
 
217
+
218
  ### Renames and Merges as a result of the Audit
219
 
220
  In several cases, it was clear from the audit that the corpora were not in the
 
231
  * `bjj` merged into the `awa` dataset
232
 
233
  ## Canaries
234
+ Canaries are provided in separate `canaries` folder. Canaries are organized into three directions: `monolingual` hosts canaries designed for the MADLAD-400 monody data, `multiway` for the multiway data, and `generic` the generic canaries generated only from the model's vocabulary.
235
 
236
  * Monolingual: Canaries here are organized by the language the canary was generated from. This corresponds exactly to the `translate_copy` setting in the paper, where the source and target language match.
237
 
 
240
  Within each subdirectory above, canaries are into separate files named by the canary type. There is always only a single file for each canary type. The `generic` folder contains within it the four canary types.
241
 
242
 
243
+ Canaries can be mixed in with normal training data to then be analyzed post-hoc to training
244
 
245
 
246
  ## References
 
258
 
259
  ## Detailed notes from the audit
260
 
261
+ Here are the notes on all languages, along with the number of documents
262
  found, and the final decision made with respect to including the language in
263
  this dataset.
264