SlimPajama-627B / README.md
rskuzma's picture
Update dataset description for right-side widget on Model Card page
2d0accd
metadata
task_categories:
  - text-generation
language:
  - en
pretty_name: SlimPajama-627B

Dataset Description

The dataset consists of 59166 jsonl files and is ~895GB compressed. It is a cleaned and deduplicated version of Together's RedPajama.

Check out our blog post explaining our methods, our code on GitHub, and join the discussion on the Cerebras Discord.

Getting Started

You can download the dataset using Hugging Face datasets:

from datasets import load_dataset
ds = load_dataset("cerebras/SlimPajama-627B")

Background

Today we are releasing SlimPajama – the largest extensively deduplicated, multi-corpora, open-source dataset for training large language models. SlimPajama was created by cleaning and deduplicating the 1.2T token RedPajama dataset from Together. By filtering out low quality data and duplicates, we were able to remove 49.6% of bytes, slimming down the dataset from 1210B to 627B tokens. We believe SlimPajama offers the highest quality and most compute efficient data to train on for runs up to 627B tokens. When upsampled, we expect SlimPajama to perform equal to or better than RedPajama-1T when training at trillion token scale.

In addition to the data, we are also releasing the tools we built to create SlimPajama. Applying MinHashLSH deduplication to trillion token datasets like RedPajama was not possible with off-the-shelf open-source code. We made several improvements to existing solutions to produce an infrastructure that can perform MinHashLSH deduplication on trillion token datasets in a distributed, multi-threaded, and memory efficient fashion. Today we are open-sourcing this infrastructure to enable the community to easily create higher quality, extensively deduplicated datasets in the future.

Our contributions

  1. SlimPajama 627B – the largest extensively deduplicated, multi-corpora, open dataset for LLM training. We release it under the Apache 2.0 license.
  2. Releasing validation and test sets, 500M tokens each, which has been decontaminated against the training data.
  3. Library of methods to replicate or pre-process from scratch other datasets. To the best of our knowledge these are the first open-source tools to enable cleaning and MinHashLSH deduplication of text data at trillion token scale.

The full set of scripts to recreate the dataset from the original RedPajama dataset are available on the Cerebras GitHub. A deeper explanation of our cleaning and deduplication process can be found in the SlimPajama blog post.

Dataset Summary

The latest research has shown that data quality is as important as data quantity. While training on more than one data epoch can be beneficial, this should be a choice rather than a side-effect of duplicates in the dataset. We decided to extensively deduplicate RedPajama to produce a dataset with higher information density. This means when using SlimPajama, you can achieve higher accuracy with the same compute budget when compared to other datasets.

Comparison of dataset features

Data source Tokens Open Source Curated Data Sources Deduplication Level
SlimPajama 627B Yes Yes Extensive
RedPajama 1.21T Yes Yes Partial
RefinedWeb-600B 600B Yes No Extensive
RefinedWeb-5T 5T No No Extensive
LLaMA 1.4T No Yes Partial
MPT 1T No Yes Partial
MassiveText 1.4T No Yes Extensive

Document low-length filter rates

Data source Document low-length filter rate
Commoncrawl 0.02%
C4 4.70%
GitHub 0.00%
Books 0.00%
ArXiv 0.62%
Wikpedia 0.00%
StackExchange 0.32%
Total 1.86%

Data source byte deduplication rates

Data source Byte deduplication rate
Commoncrawl 63.76%
C4 6.85%
GitHub 46.16%
Books 2.01%
ArXiv 0.06%
Wikipedia 2.24%
StackExchange 0.20%
Total 49.60%

Data source proportions for SlimPajama and RedPajama

Data source SlimPajama RedPajama
Commoncrawl 52.2% 72.6%
C4 26.7% 14.4%
GitHub 5.2% 4.9%
Books 4.2% 2.1%
ArXiv 4.6% 2.3%
Wikpedia 3.8% 2.0%
StackExchange 3.3% 1.7%

Languages

Primarily English, with some non-English files in Wikipedia.

Dataset Structure

The dataset consists of jsonl files, with structure as follows:

{
    "text": ...,
    "meta": {"redpajama_set_name": "RedPajamaCommonCrawl" | "RedPajamaC4" | "RedPajamaGithub" | "RedPajamaBook" | "RedPajamaArXiv" | "RedPajamaWikipedia" | "RedPajamaStackExchange"},
}

Dataset Creation

SlimPajama was created by cleaning and deduplicating the RedPajama dataset from Together via MinHashLSH. RedPajama is an open-source reproduction of the LLaMA data collection methodology.

Source Data

The data sources composing RedPajama are explained in its model card.

To cite SlimPajama, please use:

@misc{cerebras2023slimpajama,
  author = {Soboleva, Daria and Al-Khateeb, Faisal and Myers, Robert and Steeves, Jacob R and Hestness, Joel and Dey, Nolan},
  title = {{SlimPajama: A 627B token cleaned and deduplicated version of RedPajama}},
  month = June,
  year = 2023,
  howpublished = {\url{https://www.cerebras.net/blog/slimpajama-a-627b-token-cleaned-and-deduplicated-version-of-redpajama}},
  url = {https://huggingface.co/datasets/cerebras/SlimPajama-627B},
}

License

Please refer to the licenses of the data subsets you use.

Acknowledgements

  • We’d like to thank Together, Ontocord.ai, ETH DS3Lab , AAI CERC Lab for creating the original RedPajama dataset and releasing it open source.
  • This release was made possible with the support and collaboration of Opentensor.
  • Easy cloud access to Cerebras systems is provided by our partner Cirrascale.