Update Readme
#7
by
Liping
- opened
README.md
CHANGED
@@ -66,6 +66,161 @@ Each data source was filtered individually with respect to the underlying data.
|
|
66 |
## Global Deduplication
|
67 |
After the web and curated sources were filtered, all sources globally deduplicated to create TxT360. The tips and tricks behind the deduplication process are included.
|
68 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
69 |
# Citation
|
70 |
|
71 |
**BibTeX:**
|
|
|
66 |
## Global Deduplication
|
67 |
After the web and curated sources were filtered, all sources globally deduplicated to create TxT360. The tips and tricks behind the deduplication process are included.
|
68 |
|
69 |
+
## Dataset Structure
|
70 |
+
The dataset is organized under the ```data``` directory, with each subdirectory representing a data subset.
|
71 |
+
Below is an overview of the structure and organization of these subsets:
|
72 |
+
```
|
73 |
+
βββ data
|
74 |
+
βββ common-crawl # data subset
|
75 |
+
βββ CC-MAIN-2013-20 # common-crawl dumps
|
76 |
+
βββ 1-1 # number of duplicates
|
77 |
+
βββ chunk_000_0000.jsonl.gz
|
78 |
+
βββ ...
|
79 |
+
βββ 2-5
|
80 |
+
βββ chunk_000_0000.jsonl.gz
|
81 |
+
βββ ...
|
82 |
+
βββ ...
|
83 |
+
βββ CC-MAIN-2013-48
|
84 |
+
βββ 1-1
|
85 |
+
βββ chunk_000_0000.jsonl.gz
|
86 |
+
βββ ...
|
87 |
+
βββ ...
|
88 |
+
βββ ...
|
89 |
+
βββ dm_math
|
90 |
+
βββ full_data_1
|
91 |
+
βββ 0_11255.jsonl
|
92 |
+
βββ ...
|
93 |
+
βββ full_data_2
|
94 |
+
βββ 10000_11255.jsonl
|
95 |
+
βββ ...
|
96 |
+
βββ arxiv
|
97 |
+
βββ 1-1 # number of duplicates
|
98 |
+
βββ 0_171.jsonl
|
99 |
+
βββ ...
|
100 |
+
βββ 2-5
|
101 |
+
βββ 0_2.jsonl
|
102 |
+
βββ ...
|
103 |
+
βββ ...
|
104 |
+
βββ europarl
|
105 |
+
βββ 1-1 # number of duplicates
|
106 |
+
βββ 0_6.jsonl
|
107 |
+
βββ ...
|
108 |
+
βββ 2-5
|
109 |
+
βββ 0_0.jsonl
|
110 |
+
βββ ...
|
111 |
+
βββ ...
|
112 |
+
βββ ...
|
113 |
+
```
|
114 |
+
|
115 |
+
### Common Crawl (common-crawl)
|
116 |
+
Each subdirectory under ```common-crawl``` corresponds to a specific dump of the dataset.
|
117 |
+
Inside each dump folder, the data is further segmented into buckets based on the number of duplicates identified during deduplication:
|
118 |
+
|
119 |
+
- ```1-1```: Contains documents with no duplicates across the dataset.
|
120 |
+
- ```2-5```, ```6-10```, ```11-100```, ```101-1000```, ```1001-30000000```: Each contains documents that fall within the respective range of duplicates.
|
121 |
+
|
122 |
+
Example path: ```data/common-crawl/CC-MAIN-2013-20/1-1/chunk_000_0000.jsonl.gz```
|
123 |
+
|
124 |
+
### DM Math (dm_math)
|
125 |
+
The ```dm_math``` subset is divided into two subfolders to comply with the limit of 10,000 files per folder in a HuggingFace Repository:
|
126 |
+
|
127 |
+
Example path: ```data/dm_math/full_data_1/0_11255.jsonl```
|
128 |
+
|
129 |
+
### Others
|
130 |
+
Similar to common-crawl, other curated data subsets, such as arxiv, europal, etc., are organized by the number of duplicates:
|
131 |
+
- ```1-1```, ```2-5```, ```6-10```, ```11-100```, ```101-1000```, ```1001-inf```
|
132 |
+
|
133 |
+
Kindly note that some data subsets might not include the folder ```1001-inf``` (```1001-30000000``` in ```common-crawl```) or might contain only a few documents in such a folder due to the rarity of documents duplicated more than 1000 times.
|
134 |
+
|
135 |
+
## Data Schema
|
136 |
+
|
137 |
+
### Common Crawl (common-crawl)
|
138 |
+
The documents in common-crawl follow the schema:
|
139 |
+
```python
|
140 |
+
{'text': '...', # texts in the document
|
141 |
+
'meta':
|
142 |
+
{
|
143 |
+
'lang': 'en', # top 1 language detected by fastText model
|
144 |
+
'lang_score': 0.912118136882782, # language score for the detected language
|
145 |
+
'url': 'http://www.shopgirljen.com/2017/10/lg-celebrates-5-years-of-lg-oled-tv.html', # the url that raw webpage is scraped from
|
146 |
+
'timestamp': '2024-07-24T00:56:12Z', # timestamp from Common Crawl raw data
|
147 |
+
'cc-path': 'crawl-data/CC-MAIN-2024-30/segments/1720763518130.6/warc/CC-MAIN-20240723224601-20240724014601-00300.warc.gz', # the path of the document in the raw Common Crawl
|
148 |
+
'quality_signals':
|
149 |
+
{
|
150 |
+
'url_score': 0.0,
|
151 |
+
'fraction_of_duplicate_lines': 0.0,
|
152 |
+
'fraction_of_characters_in_duplicate_lines': 0.0,
|
153 |
+
'fraction_of_duplicate_paragraphs': 0.0,
|
154 |
+
'fraction_of_characters_in_duplicate_paragraphs': 0.0,
|
155 |
+
'fraction_of_characters_in_most_common_ngram': [[2, 0.03626373626373627],
|
156 |
+
[3, 0.03296703296703297],
|
157 |
+
[4, 0.01868131868131868]],
|
158 |
+
'fraction_of_characters_in_duplicate_ngrams': [[5, 0.01868131868131868],
|
159 |
+
[6, 0.01868131868131868],
|
160 |
+
[7, 0.01868131868131868],
|
161 |
+
[8, 0.0],
|
162 |
+
[9, 0.0],
|
163 |
+
[10, 0.0]],
|
164 |
+
'fraction_of_words_corrected_in_lines': 0.0,
|
165 |
+
'fraction_of_lines_ending_with_ellipsis': 0.0,
|
166 |
+
'fraction_of_lines_starting_with_bullet_point': 0.0,
|
167 |
+
'fraction_of_lines_with_toxic_words': 0.0,
|
168 |
+
'num_of_lines_with_toxic_words': 0,
|
169 |
+
'num_of_toxic_words': 0,
|
170 |
+
'word_count': 358,
|
171 |
+
'mean_word_length': 5.083798882681564,
|
172 |
+
'num_of_sentences': 19,
|
173 |
+
'symbol_to_word_ratio': 0.0,
|
174 |
+
'fraction_of_words_with_alpha_character': 1.0,
|
175 |
+
'num_of_stop_words': 82,
|
176 |
+
'num_of_paragraphs': 0,
|
177 |
+
'has_curly_bracket': False,
|
178 |
+
'has_lorem_ipsum': False,
|
179 |
+
'orig_text_has_dup_lines': False
|
180 |
+
},
|
181 |
+
'dup_signals':
|
182 |
+
{
|
183 |
+
'dup_doc_count': 166, # the number of duplicated documents
|
184 |
+
'dup_dump_count': 57, # the number of dumps that the duplicated documents are from
|
185 |
+
'dup_details': # the dump distribution of the duplicated documents
|
186 |
+
{
|
187 |
+
'2024-30': 2,
|
188 |
+
'2024-26': 1,
|
189 |
+
'2024-22': 1,
|
190 |
+
...
|
191 |
+
}
|
192 |
+
}
|
193 |
+
},
|
194 |
+
'subset': 'commoncrawl'}
|
195 |
+
```
|
196 |
+
|
197 |
+
Please note that documents without duplicates, located in folders `*/1-1/`, have an empty `dup_signals` field.
|
198 |
+
Additionally, some documents with duplicates might include an `unknown` entry within the `dup_details`.
|
199 |
+
One example could be:
|
200 |
+
```python
|
201 |
+
{'text': '...', # texts in the document
|
202 |
+
'meta':
|
203 |
+
{
|
204 |
+
...
|
205 |
+
'dup_signals':
|
206 |
+
{
|
207 |
+
'dup_doc_count': 7,
|
208 |
+
'dup_dump_count': 3,
|
209 |
+
'dup_details':
|
210 |
+
{
|
211 |
+
'unknown': 4,
|
212 |
+
'2024-30': 1,
|
213 |
+
'2024-26': 1,
|
214 |
+
'2024-22': 1,
|
215 |
+
}
|
216 |
+
}
|
217 |
+
},
|
218 |
+
'subset': 'commoncrawl'}
|
219 |
+
```
|
220 |
+
This occurs because the distribution of duplicates across dumps was not recorded in the early stages of our deduplication process, and only the total count of duplicate documents (`dup_doc_count`) was maintained.
|
221 |
+
Due to the high cost of rerunning the deduplication, we have opted to label these distributions as `unknown` when integrating them with other documents for which duplicate distribution data is available.
|
222 |
+
In these cases, the `dup_dump_count` is calculated excluding the `unknown`.
|
223 |
+
|
224 |
# Citation
|
225 |
|
226 |
**BibTeX:**
|