Chandan Akiti
commited on
Commit
·
3180859
1
Parent(s):
c705738
Update README.md
Browse files
README.md
CHANGED
@@ -1,3 +1,167 @@
|
|
1 |
---
|
2 |
-
|
|
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
task_categories:
|
3 |
+
- text-generation
|
4 |
+
language:
|
5 |
+
- en
|
6 |
+
pretty_name: Red Pajama 1T
|
7 |
---
|
8 |
+
|
9 |
+
This dataset is derived from `togethercomputer/RedPajama-Data-1T`. We removed CommonCrawl and C4 from the original RedPajama dataset
|
10 |
+
|
11 |
+
### Getting Started
|
12 |
+
|
13 |
+
The dataset consists of 2084 jsonl files.
|
14 |
+
You can download the dataset using HuggingFace:
|
15 |
+
```python
|
16 |
+
from datasets import load_dataset
|
17 |
+
ds = load_dataset("togethercomputer/RedPajama-Data-1T")
|
18 |
+
```
|
19 |
+
|
20 |
+
Or you can directly download the files using the following command:
|
21 |
+
|
22 |
+
```
|
23 |
+
wget 'https://data.together.xyz/redpajama-data-1T/v1.0.0/urls.txt'
|
24 |
+
while read line; do
|
25 |
+
dload_loc=${line#https://data.together.xyz/redpajama-data-1T/v1.0.0/}
|
26 |
+
mkdir -p $(dirname $dload_loc)
|
27 |
+
wget "$line" -O "$dload_loc"
|
28 |
+
done < urls.txt
|
29 |
+
```
|
30 |
+
|
31 |
+
|
32 |
+
A smaller 1B-token sample of the dataset can be found [here](https://huggingface.co/datasets/togethercomputer/RedPajama-Data-1T-Sample).
|
33 |
+
|
34 |
+
A full set of scripts to recreate the dataset from scratch can be found [here](https://github.com/togethercomputer/RedPajama-Data).
|
35 |
+
|
36 |
+
### Dataset Summary
|
37 |
+
|
38 |
+
RedPajama is a clean-room, fully open-source implementation of the LLaMa dataset.
|
39 |
+
|
40 |
+
| Dataset | Token Count |
|
41 |
+
|---------------|-------------|
|
42 |
+
| Commoncrawl | 878 Billion |
|
43 |
+
| C4 | 175 Billion |
|
44 |
+
| GitHub | 59 Billion |
|
45 |
+
| Books | 26 Billion |
|
46 |
+
| ArXiv | 28 Billion |
|
47 |
+
| Wikipedia | 24 Billion |
|
48 |
+
| StackExchange | 20 Billion |
|
49 |
+
| Total | 1.2 Trillion |
|
50 |
+
|
51 |
+
### Languages
|
52 |
+
|
53 |
+
Primarily English, though the Wikipedia slice contains multiple languages.
|
54 |
+
|
55 |
+
## Dataset Structure
|
56 |
+
|
57 |
+
The dataset structure is as follows:
|
58 |
+
|
59 |
+
```json
|
60 |
+
{
|
61 |
+
"text": ...,
|
62 |
+
"meta": {"url": "...", "timestamp": "...", "source": "...", "language": "...", ...},
|
63 |
+
"red_pajama_subset": "common_crawl" | "c4" | "github" | "books" | "arxiv" | "wikipedia" | "stackexchange"
|
64 |
+
}
|
65 |
+
```
|
66 |
+
|
67 |
+
## Dataset Creation
|
68 |
+
|
69 |
+
This dataset was created to follow the LLaMa paper as closely as possible to try to reproduce its recipe.
|
70 |
+
|
71 |
+
### Source Data
|
72 |
+
|
73 |
+
#### Commoncrawl
|
74 |
+
|
75 |
+
We download five dumps from Commoncrawl, and run the dumps through the official `cc_net` pipeline.
|
76 |
+
We then deduplicate on the paragraph level, and filter out low quality text using a linear classifier trained to
|
77 |
+
classify paragraphs as Wikipedia references or random Commoncrawl samples.
|
78 |
+
|
79 |
+
#### C4
|
80 |
+
|
81 |
+
C4 is downloaded from Huggingface. The only preprocessing step is to bring the data into our own format.
|
82 |
+
|
83 |
+
#### GitHub
|
84 |
+
|
85 |
+
The raw GitHub data is downloaded from Google BigQuery. We deduplicate on the file level and filter out low quality
|
86 |
+
files and only keep projects that are distributed under the MIT, BSD, or Apache license.
|
87 |
+
|
88 |
+
#### Wikipedia
|
89 |
+
We use the Wikipedia dataset available on Huggingface, which is based on the Wikipedia dump from 2023-03-20 and contains
|
90 |
+
text in 20 different languages. The dataset comes in preprocessed format, so that hyperlinks, comments and other
|
91 |
+
formatting boilerplate has been removed.
|
92 |
+
|
93 |
+
#### Gutenberg and Books3
|
94 |
+
The PG19 subset of the Gutenberg Project and Books3 datasets are downloaded from Huggingface. After downloading, we use
|
95 |
+
simhash to remove near duplicates.
|
96 |
+
|
97 |
+
#### ArXiv
|
98 |
+
ArXiv data is downloaded from Amazon S3 in the `arxiv` requester pays bucket. We only keep latex source files and
|
99 |
+
remove preambles, comments, macros and bibliographies.
|
100 |
+
|
101 |
+
#### Stackexchange
|
102 |
+
The Stack Exchange split of the dataset is download from the
|
103 |
+
[Internet Archive](https://archive.org/download/stackexchange). Here we only keep the posts from the 28 largest sites,
|
104 |
+
remove html tags, group the posts into question-answer pairs, and order answers by their score.
|
105 |
+
|
106 |
+
### SHA256 Checksums
|
107 |
+
|
108 |
+
SHA256 checksums for the dataset files for each data source are available here:
|
109 |
+
|
110 |
+
```
|
111 |
+
https://data.together.xyz/redpajama-data-1T/v1.0.0/sha256/arxiv_SHA256SUMS.txt
|
112 |
+
https://data.together.xyz/redpajama-data-1T/v1.0.0/sha256/book_SHA256SUMS.txt
|
113 |
+
https://data.together.xyz/redpajama-data-1T/v1.0.0/sha256/c4_SHA256SUMS.txt
|
114 |
+
https://data.together.xyz/redpajama-data-1T/v1.0.0/sha256/common_crawl_SHA256SUMS.txt
|
115 |
+
https://data.together.xyz/redpajama-data-1T/v1.0.0/sha256/github_SHA256SUMS.txt
|
116 |
+
https://data.together.xyz/redpajama-data-1T/v1.0.0/sha256/stackexchange_SHA256SUMS.txt
|
117 |
+
https://data.together.xyz/redpajama-data-1T/v1.0.0/sha256/wikipedia_SHA256SUMS.txt
|
118 |
+
```
|
119 |
+
|
120 |
+
To cite RedPajama, please use:
|
121 |
+
|
122 |
+
```
|
123 |
+
@software{together2023redpajama,
|
124 |
+
author = {Together Computer},
|
125 |
+
title = {RedPajama: An Open Source Recipe to Reproduce LLaMA training dataset},
|
126 |
+
month = April,
|
127 |
+
year = 2023,
|
128 |
+
url = {https://github.com/togethercomputer/RedPajama-Data}
|
129 |
+
}
|
130 |
+
```
|
131 |
+
|
132 |
+
### License
|
133 |
+
Please refer to the licenses of the data subsets you use.
|
134 |
+
|
135 |
+
* [Common Crawl Foundation Terms of Use](https://commoncrawl.org/terms-of-use/full/)
|
136 |
+
* [C4 license](https://huggingface.co/datasets/allenai/c4#license)
|
137 |
+
* GitHub was limited to MIT, BSD, or Apache licenses only
|
138 |
+
* Books: [the_pile_books3 license](https://huggingface.co/datasets/the_pile_books3#licensing-information) and [pg19 license](https://huggingface.co/datasets/pg19#licensing-information)
|
139 |
+
* [ArXiv Terms of Use](https://info.arxiv.org/help/api/tou.html)
|
140 |
+
* [Wikipedia License](https://huggingface.co/datasets/wikipedia#licensing-information)
|
141 |
+
* [StackExchange license on the Internet Archive](https://archive.org/details/stackexchange)
|
142 |
+
|
143 |
+
<!--
|
144 |
+
### Annotations
|
145 |
+
#### Annotation process
|
146 |
+
[More Information Needed]
|
147 |
+
#### Who are the annotators?
|
148 |
+
[More Information Needed]
|
149 |
+
### Personal and Sensitive Information
|
150 |
+
[More Information Needed]
|
151 |
+
## Considerations for Using the Data
|
152 |
+
### Social Impact of Dataset
|
153 |
+
[More Information Needed]
|
154 |
+
### Discussion of Biases
|
155 |
+
[More Information Needed]
|
156 |
+
### Other Known Limitations
|
157 |
+
[More Information Needed]
|
158 |
+
## Additional Information
|
159 |
+
### Dataset Curators
|
160 |
+
[More Information Needed]
|
161 |
+
### Licensing Information
|
162 |
+
[More Information Needed]
|
163 |
+
### Citation Information
|
164 |
+
[More Information Needed]
|
165 |
+
### Contributions
|
166 |
+
[More Information Needed]
|
167 |
+
-->
|