RedStone
RedStone is an innovative and scalable pipeline designed to extract and process data from a vast amount of web content, facilitating the creation of diverse and comprehensive pre-training datasets. See the repo for more details.
Description
Since we do not have the permission to open-source the processed data, this repository contains an index of high-quality pages as determined by RedStone-Web after filtering. Using this index, one can easily extract pages deemed high-quality by RedStone from the original Common Crawl dataset. Anyone can use the code provided by RedStone to replicate RedStone-Web, and can also apply their own processing code for further refinement.
If you have the appropriate licenses, we encourage you to use these scripts to reproduce the dataset and contribute it to the open-source community. We will reference the data here for easy access. Additionally, we welcome you to use RedStone to expand domain-specific categories beyond just code, math, and QA.
By using this index and processing script, we are ultimately able to achieve approximately 3.17 trillion tokens pretraining dataset.
Schema
uri
:WARC-Target-URI
in the original WARC/WET filecc_snapshot
: Shorter version of CommonCrawl's snapshot id (e.g. CC-MAIN-2021-49 -> 202149)cc_path
: URL to the source CommonCrawl filerecord_id
(WARC only):WARC-Record-ID
in the original WARC filedigest
(WET only):WARC-Block-Digest
in the original WET file
Reproduce
WARC
- Please reference our code repo
WET
- Line-level deduplication
- This is the first step of WET based data processing and should be done before filtering. In our experience, this step requires at least a snapshot of the raw CC WET data to achieve good results. Due to the data scale, we recommended running this step with a cluster computing system like Spark.
- Index matching & language identification
- After the initial line-level deduplication, you can apply our filtered index to your data
- Minhash deduplication
- Another deduplcation step that needs cluster computing. You can reference our repo for our Minhash-LSH related code and a sample Spark deduplication script
Responsible AI FAQ
- What is RedStone Source Code?
- RedStone is a pipeline designed to extract a wide range of specified knowledge from Common Crawl on a large scale. It is composed of three modules, Collection, Filtering and Extraction. As an example, we use RedStone to build extensive domain-specific datasets in the fields of code, mathematics, question answering (QA), and general data. Utilizing RedStone, it is possible to easily acquire valuable knowledge from a multitude of other domains within Common Crawl.
- What can RedStone Source Code do?
- RedStone Source Code provides the sample codes of the pipeline’s components, workflow and index of source location, enabling anyone to construct large-scale various domains from Common Crawl, including general web content, web code, web mathematics and web QA data.
- What is/are RedStone Source Code’s intended use(s)?
- We release RedStone, aiming to provide this resource to the research community to accelerate the development of large language models and for demonstrating a novel method of constructing training datasets. Given the research nature of this work, production or commercial uses are out of scope without further testing and mitigation.
- How was RedStone Source Code evaluated? What metrics are used to measure performance?
- We use RedStone to build domain-specific datasets in the fields of code, mathematics, question answering (QA), and general datasets as examples. We evaluate the performance of the datasets across multiple benchmarks, demonstrating that RedStone significantly enhances model performance in mathematics, code, and QA tasks.
- What are the limitations of [RedStone Source Code]? How can users minimize the impact of RedStone dataset’s limitations when using the system?
- RedStone takes several domains as examples to verify the methodologies and pipelines. We believe the ways should work for other fields. However, the source code repo is customized for these domain and English materials only. It takes extra effort to revise the codes for your tasks and setting if you would like to obtain data of different domain, languages with your environment.
- RedStone employs quality filters to get content with correct grammar, logical consistency, and factual accuracy. Despite our efforts to remove toxic content, some harmful content may be present.
- RedStone used scope of deduplication, which indicates that narrowing the scope of deduplication yields the highest scores. A possible explanation is that a narrower deduplication scope results in a data distribution that more closely mirrors the real world, where frequently occurring data in real life also appears multiple times in the dataset. However, we are currently unable to verify this hypothesis and will investigate it.
- There might be incorrect data in raw data that could not be filtered out, which may result in inaccurate answers for some questions.
- Common Crawl data may not be suitable for all downstream uses due to copyright or other legal reasons. Users are responsible for verifying the legal right to use Common Crawl data for their intended purpose.
- What operational factors and settings allow for effective and responsible use of RedStone Source Code?
- The user is responsible for validating the safety and accuracy of any datasets developed using RedStone Source Code, or any model developed using a dataset constructed using our methods.
Citation
If you find this repository useful, please consider citing our work:
@misc{chang2024redstonecuratinggeneralcode,
title={RedStone: Curating General, Code, Math, and QA Data for Large Language Models},
author={Yaoyao Chang and Lei Cui and Li Dong and Shaohan Huang and Yangyu Huang and Yupan Huang and Scarlett Li and Tengchao Lv and Shuming Ma and Qinzheng Sun and Wenhui Wang and Furu Wei and Ying Xin and Mao Yang and Qiufeng Yin and Xingxing Zhang},
year={2024},
eprint={2412.03398},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2412.03398},
}
License
The content of this project itself is licensed under the MIT
Microsoft Open Source Code of Conduct
Contact
For help or issues using RedStone, please submit a GitHub issue.
For other communications related to RedStone, please contact Lei Cui or Furu Wei.
- Downloads last month
- 231