amphora commited on
Commit
efbac1f
1 Parent(s): 966405d

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +80 -0
README.md CHANGED
@@ -21,3 +21,83 @@ configs:
21
  - split: train
22
  path: data/train-*
23
  ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
21
  - split: train
22
  path: data/train-*
23
  ---
24
+
25
+ # KOREAN-WEBTEXT-1B
26
+
27
+ **KOREAN-WEBTEXT-1B** is a high-quality Korean language corpus consisting of 1.04 billion tokens. The data has been collected from the following sources:
28
+
29
+ - **cc100**
30
+ - **oscar-corpus/OSCAR-2201**
31
+ - **ontocord/CulturaY**
32
+ - **Additional credible internet sources collected by out team**
33
+
34
+ (We are working to add more sources)
35
+
36
+ The dataset undergoes rigorous filtering at both the sentence and document levels to ensure quality of text data. Additionally, simple deduplication processes are applied to further refine the dataset.
37
+
38
+ ## Dataset Structure
39
+
40
+ ### Sentence-Level Filters
41
+
42
+ The following filters are applied at the sentence level:
43
+
44
+ 1. **Repetition Check**: The ratio of repetition for any word in a line should not exceed 0.2.
45
+ 2. **Punctuation Check**: Lines must end with one of these punctuation marks: `.`, `?`, `]`, or `"`.
46
+ 3. **Token Count Check**: The line must contain more than 16 tokens.
47
+ 4. **Character Count Check**: The line must contain more than 32 characters.
48
+
49
+ ### Document-Level Filters
50
+
51
+ The following filters are applied at the document level:
52
+
53
+ 1. **Token Count Check**: Documents must contain more than 512 tokens.
54
+ 2. **Stopwords Removal**: Documents containing any of the following stopwords are removed:
55
+ ```python
56
+ stopwords = [
57
+ 'www', 'http', '...', 'ㅋㅋㅋ', '약관', 'is', '카지노', '토토', '\u3000',
58
+ '■', '▲', '010', '.kr', '@', '마사지', '스웨디시', '대선'
59
+ ]
60
+ ```
61
+
62
+ ### Deduplication Processes
63
+
64
+ To ensure data uniqueness, the following deduplication steps are applied:
65
+
66
+ 1. **Exact Deduplication**: Removal of exact duplicate lines.
67
+ 2. **First 15 Tokens Deduplication**: Removal of lines with identical first 15 tokens.
68
+ 3. **Last 15 Tokens Deduplication**: Removal of lines with identical last 15 tokens.
69
+
70
+ ## Usage
71
+
72
+ While the dataset may be small for pretraining models due to its size, we expect it to be better suited for ablation studies.
73
+
74
+ ### Examples
75
+
76
+ #### Loading the Dataset
77
+
78
+ To load and use the dataset, you can use the following example code:
79
+
80
+ ```python
81
+ import datasets
82
+
83
+ dataset = datasets.load_dataset('HAERAE-HUB/KOREAN-WEBTEXT-1B')
84
+ ```
85
+
86
+ ## Citation
87
+
88
+ If you use this dataset in your research, please cite it as follows:
89
+
90
+ ```
91
+ @dataset{KOREAN-WEBTEXT-1B,
92
+ title={KOREAN-WEBTEXT-1B: A High-Quality Korean Language Corpus},
93
+ author={HAERAE-Team},
94
+ year={2024},
95
+ howpublished={\url{https://huggingface.co/datasets/HAERAE-HUB/KOREAN-WEBTEXT-1B}},
96
+ }
97
+ ```
98
+
99
+ ## Contact
100
+
101
+ For more information or questions about the dataset, please contact the maintainers at [spthsrbwls123@yonsei.ac.kr].
102
+
103
+ ---