This repository contains the code for the RedPajama-V2 dataset. For more information on the dataset, check out our blog post. The dataset is also available on HuggingFace. For the code used for the RedPajama-1T dataset, please refer to the rp_v1 branch in this repo.

RedPajama-V2 is an open dataset for training large language models. The dataset includes over 100B text documents coming from 84 CommonCrawl snapshots and processed using the CCNet pipeline. Out of these, there are 30B documents in the corpus that additionally come with quality signals, and 20B documents that are deduplicated.

The number of documents and tokens for the annotated and deduplicated head_middle part of the dataset is shown in the table below.

English, German, French, Italian, Spanish

Copy the file configs/rp_v2.0.conf to e.g. configs/default.conf and configure the environment variables. These will be used throughout the pipeline.

To run with docker, build the docker image using

. configs/default.conf
cd app
docker build -t "${DOCKER_REPO}:" .

Also, make sure you have s5cmd installed and your S3 profile configured so that you can pull data from an S3 bucket.

You can run the steps of the pipeline without any containerized environment. However, the running scripts assume you have a docker and apptainer installation.

The pipeline is composed of three steps, namely 1) preparing artifacts, 2) computing quality signals, and 3) deduplication.

Important: In case you are not running steps (1) and (2) with the provided scripts (i.e., docker containers built with the provided Dockerfile), make sure to set the PYTHONHASHSEED environment variable to a consistent value (e.g., 42) using

export PYTHONHASHSEED=42

This is to ensure consistency of hash functions used in the computation of DSIR weights.

This part of the pipeline creates the artifacts that are used in the subsequent steps. This includes building quality classifiers, training bag-of-ngram generative models for importance weight computation, fetching the list of bad words from the LDNOOBW repo, and fetching the most recent list of blacklisted urls from the UT1 blacklist.

As a first step, download the english wikipedia reference classifier from here and place it in ${DATA_ROOT}/wikiref-model/en/en-model.bin. This is the same fasttext classifier that was used in RedPajama-V1.

To create the remaining artifacts, make sure that the environment variables are set in the config file. Then, from the root directory of the repository, run

bash scripts/run_prep_artifacts.sh \
  --config configs/rp_v2.0.conf \
  --listings /path/to/listings/file.txt\
  --max_workers 32

where /path/to/listings/file.txt is a file that contains the keys to the ccnet data that you want to process (e.g., 2023-06/0000/en_head.json.gz).

You can set the max_workers flag to the number of parallel processes you want to use.

This step will generate an id which you can store in the environment variable ARTIFACTS_ID for the next step.

The second step of the pipeline compute the quality signals, including the minhash signatures to run fuzzy deduplication in the subsequent step. To run this step, make sure the environment variables are set in the config file. Then, from the root directory of the repository, run

bash scripts/apptainer_run_quality_signals.sh \
  --config configs/rp_v2.0.conf \
  --dump_id "2022-49" \
  --input_base_uri "file:///path/to/data/root" \
  --output_base_uri "file:///path/to/outout/data/root" \
  --max_docs -1

The third component of the pipeline consists of deduplication steps. Here we provide code to run exact and fuzzy deduplication.

Content based deduplication is implemented in app/src/bloomfilter.py. It can be run independently of the previous step, but the data needs to stored in an S3 bucket. For this step, from the app directory, run:

python3 app/src/bloomfilter.py \
  --listings /path/to/listings/file.txt \
  --input_base_uri "s3://path/to/ccnet/data" \
  --output_dir "/path/to/output" \
  --s3_profile "..." \
  --endpoint_url "..." \
  --parallel_readers 32 \
  --batch_size 10 \
  --capacity "..." \
  --error_rate "..."

It is important to choose the correct capacity (i.e., > #documents), since otherwise the error_rate will not be guaranteed and more false positives will appear. The implementation is based on the pybloomfiltermmap3 library.

In the third step of the pipeline, we run locality sensitive hashing on the minhash signatures generated in the first step. To run this step, make sure that you use the same configuration as in the quality signals step. Then, from the root directory of the repository, run

bash scripts/apptainer_run_lsh.sh \
  --config configs/rp_v2.0.conf \
  --dump_id "2022-49" \
  --input_base_uri "file:///path/to/data/root" \
  --output_dir "/path/to/output" \
  --similarity "<similarity_threshold>" \
  --listings "/minhash/listings/file.txt" \
  --max_docs -1

The implementation is based on polars and was tested with 200M documents on a 64 core machine with 500G of RAM.

The second step of this pipeline computes the following set of quality signals. We hope to grow this list further over time as more signals are developed.

We are appreciative to so many partners and collaborators that together are pushing forward the frontier of open LLM models.

• Thank you to the OLMo team at AI2 and friends at OpenGPT-X for the insightful discussions about datasets and data quality! Also for everyone who builds on the RedPajama dataset, including Cerebras for their SlimPajama efforts, and the over 500 models built on RedPajam to date by the open-source AI community.
• We are grateful to the great team at EleutherAI for paving the path on open training datasets with The Pile and for open-sourcing code we use in training some of the RedPajama models.
• Thank you to our partners of RedPajama-v1, including Ontocord.ai, MILA Québec AI Institute, ETH DS3Lab, Université de Montréal, Stanford Center for Research on Foundation Models (CRFM), Stanford Hazy Research research group and LAION.

Copyright 2023 Together Computer

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

   http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.

For full terms, see the LICENSE file. If you have any questions, comments, or concerns about licensing please contact us.

For the dataset itself, please refer to the Common Crawl Foundation Terms of Use.

To cite RedPajama, please use:

@software{together2023redpajama,
  author = {Together Computer},
  title = {RedPajama: an Open Dataset for Training Large Language Models},
  month = October,
  year = 2023,
  url = {https://github.com/togethercomputer/RedPajama-Data}
}