zhangce commited on
Commit
753f485
1 Parent(s): 93f12f7

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +108 -0
README.md ADDED
@@ -0,0 +1,108 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ### Dataset Summary
2
+
3
+ RedPajama is a clean-room, fully open-source implementation of the LLaMa dataset.
4
+
5
+ | Dataset | Token Count |
6
+ |---------------|-------------|
7
+ | Commoncrawl | 878 Billion |
8
+ | C4 | 175 Billion |
9
+ | GitHub | 59 Billion |
10
+ | Books | 26 Billion |
11
+ | ArXiv | 28 Billion |
12
+ | Wikipedia | 24 Billion |
13
+ | StackExchange | 20 Billion |
14
+ | Total | 1.2 Trillion |
15
+
16
+ The dataset consists of 2084 jsonl files, whose URLs are listed in the following file:
17
+ ```
18
+ https://data.together.xyz/redpajama-data-1T/v1.0.0/urls.txt
19
+ ```
20
+
21
+ To download the data, once can use the following command:
22
+ ```
23
+ wget -i https://data.together.xyz/redpajama-data-1T/v1.0.0/urls.txt
24
+ ```
25
+
26
+ A smaller 1B-token sample of the dataset can be found [here](https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T-Sample).
27
+
28
+ A full set of scripts to recreate the dataset from scratch can be found [here](https://github.com/togethercomputer/RedPajama-Data).
29
+
30
+ ### Languages
31
+
32
+ Primarily English, though the Wikipedia slice contains multiple languages.
33
+
34
+ ## Dataset Structure
35
+
36
+ The dataset structure is as follows:
37
+
38
+ ```
39
+ {
40
+ "text": ...,
41
+ "meta": {"url": "...", "timestamp": "...", "source": "...", "language": "...", ...}
42
+ }
43
+ ```
44
+
45
+ ## Dataset Creation
46
+
47
+ This dataset was created to follow the LLaMa paper as closely as possible to try to reproduce its recipe.
48
+
49
+ ### Source Data
50
+
51
+ #### Commoncrawl
52
+
53
+ We download five dumps from Commoncrawl, and run the dumps through the official `cc_net` pipeline.
54
+ We then deduplicate on the paragraph level, and filter out low quality text using a linear classifier trained to
55
+ classify paragraphs as Wikipedia references or random Commoncrawl samples.
56
+
57
+ #### C4
58
+
59
+ C4 is downloaded from Huggingface. The only preprocessing step is to bring the data into our own format.
60
+
61
+ #### GitHub
62
+
63
+ The raw GitHub data is downloaded from Google BigQuery. We deduplicate on the file level and filter out low quality
64
+ files and only keep projects that are distributed under the MIT, BSD, or Apache license.
65
+
66
+ #### Wikipedia
67
+ We use the Wikipedia dataset available on Huggingface, which is based on the Wikipedia dump from 2023-03-20 and contains
68
+ text in 20 different languages. The dataset comes in preprocessed format, so that hyperlinks, comments and other
69
+ formatting boilerplate has been removed.
70
+
71
+ #### Gutenberg and Books3
72
+ The PG19 subset of the Gutenberg Project and Books3 datasets are downloaded from Huggingface. After downloading, we use
73
+ simhash to remove near duplicates.
74
+
75
+ #### ArXiv
76
+ ArXiv data is downloaded from Amazon S3 in the `arxiv` requester pays bucket. We only keep latex source files and
77
+ remove preambles, comments, macros and bibliographies.
78
+
79
+ #### Stackexchange
80
+ The Stack Exchange split of the dataset is download from the
81
+ [Internet Archive](https://archive.org/download/stackexchange). Here we only keep the posts from the 28 largest sites,
82
+ remove html tags, group the posts into question-answer pairs, and order answers by their score.
83
+
84
+ <!--
85
+ ### Annotations
86
+ #### Annotation process
87
+ [More Information Needed]
88
+ #### Who are the annotators?
89
+ [More Information Needed]
90
+ ### Personal and Sensitive Information
91
+ [More Information Needed]
92
+ ## Considerations for Using the Data
93
+ ### Social Impact of Dataset
94
+ [More Information Needed]
95
+ ### Discussion of Biases
96
+ [More Information Needed]
97
+ ### Other Known Limitations
98
+ [More Information Needed]
99
+ ## Additional Information
100
+ ### Dataset Curators
101
+ [More Information Needed]
102
+ ### Licensing Information
103
+ [More Information Needed]
104
+ ### Citation Information
105
+ [More Information Needed]
106
+ ### Contributions
107
+ [More Information Needed]
108
+ -->