Home

Awesome

DataComp-LM (DCLM)

Table of Contents

Introduction

DataComp-LM (DCLM) is a comprehensive framework designed for building and training large language models (LLMs) with diverse datasets. It offers a standardized corpus of over 300T unfiltered tokens from CommonCrawl, effective pretraining recipes based on the open_lm framework, and an extensive suite of over 50 evaluations. This repository provides tools and guidelines for processing raw data, tokenizing, shuffling, training models, and evaluating their performance.

DCLM enables researchers to experiment with various dataset construction strategies across different compute scales, from 411M to 7B parameter models. Our baseline experiments show significant improvements in model performance through optimized dataset design.

Already, DCLM has enabled the creation of several high quality datasets that perform well across scales and outperform all open datasets. Accuracy vs compute tradeoff

<p align="center"> <em><b>Developing datasets for better models that are cheaper to train.</b> Using DataComp-LM, we develop a high-quality dataset, DCLM-BASELINE, which we use to train models with strong compute performance tradeoffs. We compare on both a Core set of tasks (left) and on MMLU 5-shot (right). DCLM-BASELINE (orange) shows favorable performance relative to both close-source models (crosses) and other open-source datasets and models (circles).</em> </p>

Submission workflow:

For more details, please refer to our paper.

Leaderboard

The DCLM leaderboard showcases the performance of models trained on various scales and datasets. The leaderboard is updated regularly with the latest submissions from the community.

Below are comparisions of our model with others in the 7B regime.

ModelParamsTokensOpen dataset?COREMMLUEXTENDED
Open weights, closed datasets
Llama27B2T49.245.834.1
DeepSeek7B2T50.748.535.3
Mistral-0.37B?57.062.745.1
QWEN-27B?57.571.950.5
Llama38B15T57.666.246.3
Gemma8B6T57.864.344.6
Phi-37B?61.069.957.9
Open weights, open datasets
Falcon7B1T44.127.425.1
OLMo-1.77B2.1T47.054.034.2
MAP-Neo7B4.5T50.257.140.4
Models we trained
FineWeb edu7B0.14T38.726.322.1
FineWeb edu7B0.28T41.937.324.5
DCLM-BASELINE7B0.14T44.138.325.0
DCLM-BASELINE7B0.28T48.950.831.8
DCLM-BASELINE7B2.6T57.163.745.4

Getting Started

To get started with DCLM, follow these steps:

  1. Clone the repository:

    git clone https://github.com/mlfoundations/DCLM.git
    cd DCLM
    
  2. Install dependencies:

    pip install -r requirements.txt
    

    Before installing the dependencies, make sure cmake, build-essential, and g++ are installed, e.g., by installing:

    apt install cmake build-essential
    apt install g++-9
    update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-9 90
    

    To download additional models and data needed for baseline reproduction, run:

    python setup.py install
    
  3. Set up your environment: DCLM uses AWS for storage and possible as a compute backend, and ray for distributed processing. Ensure you have the necessary environment variables and configurations for AWS and Ray clusters.

    We recommend the use of Python 3.10 with DCLM.

Selecting Raw Sources

If you are creating a new source:

If you are selecting a raw source for downstream processing:

Processing the Data

To process raw data, follow these steps:

  1. Define a set of processing steps: Create a pipeline config YAML file specifying the operations. See our reproduction of C4 for example. Further details on defining a pipeline can be found here.

  2. Set up a Ray cluster: The data processing script relies on Ray for distributed processing of data. This cluster can be either launched on a single node (for small scale data processing) or using AWS EC2 instances.

    To launch a local cluster, use the following command:

    ray start --head --port 6379
    

    To launch a cluster using AWS EC2 instances, use the following:

    ray up <your_cluster_config>
    

    where <your_cluster_config> is a cluster configuration script that depends on your specific use case. We invite the reader to go over the Ray documentation for details on how to create this config file.

    Important: When using EC2 instances, make sure to tear down your cluster after your job finishes, so as to not incur unnecessary costs!

    A sample config file can be seen here (make sure to adapt to your needs):

    cluster_name: test-processing
    max_workers: 2
    upscaling_speed: 1.0
    available_node_types:
        ray.head.default:
            resources: {}
            node_config:
                ImageId: ami-0c5cce1d70efb41f5
                InstanceType: i4i.4xlarge
                IamInstanceProfile:
                    # Replace 000000000000 with your IAM account 12-digit ID
                    Arn: arn:aws:iam::000000000000:instance-profile/ray-autoscaler-v1
        ray.worker.default:
            min_workers: 2
            max_workers: 2
            node_config:
                ImageId: ami-0c5cce1d70efb41f5
                InstanceType: i4i.4xlarge
                IamInstanceProfile:
                    # Replace 000000000000 with your IAM account 12-digit ID
                    Arn: arn:aws:iam::000000000000:instance-profile/ray-autoscaler-v1
    
    # Cloud-provider specific configuration.
    provider:
        type: aws
        region: us-west-2
        cache_stopped_nodes: False
    
    setup_commands:
        - sudo mkfs -t xfs /dev/nvme1n1
        - sudo mount /dev/nvme1n1 /tmp
        - sudo chown -R $USER /tmp
        - sudo chmod -R 777 /tmp
        - wget https://repo.anaconda.com/miniconda/Miniconda3-py310_23.3.1-0-Linux-x86_64.sh -O miniconda.sh
        - bash ~/miniconda.sh -f -b -p /tmp/miniconda3/
        - echo 'export PATH="/tmp/miniconda3/bin/:$PATH"' >> ~/.bashrc
        # Include your AWS CREDS here
        - echo 'export AWS_ACCESS_KEY_ID=' >> ~/.bashrc
        - echo 'export AWS_SECRET_ACCESS_KEY=' >> ~/.bashrc
        - pip install --upgrade pip setuptools wheel
        - pip install -U "ray[default] @ https://s3-us-west-2.amazonaws.com/ray-wheels/latest/ray-3.0.0.dev0-cp310-cp310-manylinux2014_x86_64.whl"
        - pip install boto3==1.26.90
        - pip install s3fs==2022.11.0
        - pip install psutil
        - pip install pysimdjson
        - pip install pyarrow
        - git clone https://github.com/mlfoundations/dclm.git
        - pip install -r dclm/requirements.txt
        - cd dclm && python3 setup.py install
    
  3. Run the processing script: To run the processing script, in the case of a local cluster, simply run the following command:

    python3 ray_processing/process.py --source_ref_paths <source_json> --readable_name <name> --output_dir <s3_output_dir> --config_path <config_yaml> --source_name <source_name>
    

    When using EC2 instances, you need to connect to the cluster and then launch the command

    # In your local terminal
    ray attach <your_cluster_config>
    
    # Inside the cluster EC2 instance
    cd dcnlp
    export PYTHONPATH=$(pwd)
    python3 ray_processing/process.py --source_ref_paths <source_json> --readable_name <name> --output_dir <s3_output_dir> --config_path <config_yaml> --source_name <source_name>
    
  4. Monitor and tear down: You can track the progress of data processing via the global_stats.jsonl file in the output directory. After the job finishes, you can tear down your cluster via ray stop (in the local cluster case) or ray down <your_cluster_config> (in the AWS EC2 case). THIS IS VERY IMPORTANT TO NOT INCUR ADDITIONAL COSTS WHEN USING EC2!

Deduplication

To deduplicate the raw text as we have done in DCLM-Baseline, use the tools provided in the dedup subdirectory. Here we include several rust tools for deduplication, but we recommend using BFF, located in dedup/bff. Specific instructions to run deduplication are contained in the readme in each of the directories containing the rust tools.

We note that the code in dedup specifically refers to inter-document fuzzy deduplication, i.e., identifying near-duplicates across documents in the corpus. Tooling built in Ray to identify exact content and URL duplicates is contained in ray_processing/dedup_jsonl.py (but we do not use this form of dedup in DCLM-Baseline).

Tokenize and Shuffle

After processing the raw text, you should convert it into tokenized datasets and perform shuffling for training:

  1. Set up a Ray cluster: Set up a Ray cluster in the same way as the processing step.

  2. Run the tokenize and shuffle script:

    python ray_processing/tokenize_shuffle.py --source_ref_paths <source_jsons> --readable_name <name> --output <s3_output_dir> --content_key text --do_sample --default_dataset_yaml <mixing_yaml>
    
  3. Tear down: Tear down the Ray cluster as in the processing step.

The tokenize_shuffle.py script creates a dataset in webdataset format, along with a manifest.jsonl file. This file is required by the training script, and it contains information on the number of sequences inside each shard of the dataset. If needed, this manifest file can also be created manually, via the following command:

python -m open_lm.utils.make_wds_manifest --data-dir <tokenized_data_dir>

Model Training

To train a model using the tokenized dataset:

  1. Run the training script:
    torchrun --nproc-per-node 8 -m training.train --scale <scale> <tokenized_json> --logs <log_dir> [--remote-sync <s3_bucket>] [--chinchilla-multiplier <multiplier>] [--clean-exp] [--report-to-wandb]
    

You can expect the following training times per track:

ScaleModel parametersTrain tokensTrain FLOPsTrain H100 hoursPool size
400M-1x412M8.2B2.0e1926137B
1B-1x1.4B28B2.4e202401.64T
1B-5x1.4B138B1.2e2112008.20T
7B-1x6.9B138B5.7e2137007.85T
7B-2x6.9B276B1.1e22730015.7T
  1. Monitor and manage your training jobs: Use slurm sbatch scripts or Sagemaker for running experiments on various compute infrastructures.

Evaluation

Evaluate trained models using the following methods:

  1. Preferred Method:

    python tools/eval_expdb.py --start_idx 0 --end_idx 3 --filters name=<filter> --prefix_replacement <prefix_replacement> --num_gpus 8 --output_dir <s3_output_dir> --eval_yaml <eval_yaml>
    
  2. Direct Evaluation:

    torchrun --nproc_per_node <num_gpus> eval/eval_openlm_ckpt.py --checkpoint <checkpoint> --eval-yaml <eval_yaml> --config <model_params_file> --model <open_lm_config> --output-file <output_file_path>
    

Submission

When you finished training and evaluating your model, a model eval json file has been generated and is at exp_data/evals. You can now open a pull request to the main repository to share your results with the team and submit it to the leaderboard.

Contributing

We welcome contributions to improve the DCLM framework. Please follow our contributing guide for submitting pull requests and reporting issues.

Downloading Artifacts

Datasets

We provide multiple datasets, both as starting points for each of the competition scales, as well as the results of our processing pipeline.

Pretrained Models

We provide links to models pretrained using our dataset via the DCLM collection on Huggingface, found here. These models can be downloaded and evaluated using the OpenLM library.

How to Cite Us

If you use our dataset or models in your research, please cite us as follows:

@article{li2024datacomplm,
      title={DataComp-LM: In search of the next generation of training sets for language models}, 
      author={Jeffrey Li and Alex Fang and Georgios Smyrnis and Maor Ivgi and Matt Jordan and Samir Gadre and Hritik Bansal and Etash Guha and Sedrick Keh and Kushal Arora and Saurabh Garg and Rui Xin and Niklas Muennighoff and Reinhard Heckel and Jean Mercat and Mayee Chen and Suchin Gururangan and Mitchell Wortsman and Alon Albalak and Yonatan Bitton and Marianna Nezhurina and Amro Abbas and Cheng-Yu Hsieh and Dhruba Ghosh and Josh Gardner and Maciej Kilian and Hanlin Zhang and Rulin Shao and Sarah Pratt and Sunny Sanyal and Gabriel Ilharco and Giannis Daras and Kalyani Marathe and Aaron Gokaslan and Jieyu Zhang and Khyathi Chandu and Thao Nguyen and Igor Vasiljevic and Sham Kakade and Shuran Song and Sujay Sanghavi and Fartash Faghri and Sewoong Oh and Luke Zettlemoyer and Kyle Lo and Alaaeldin El-Nouby and Hadi Pouransari and Alexander Toshev and Stephanie Wang and Dirk Groeneveld and Luca Soldaini and Pang Wei Koh and Jenia Jitsev and Thomas Kollar and Alexandros G. Dimakis and Yair Carmon and Achal Dave and Ludwig Schmidt and Vaishaal Shankar},
      year={2024},
      journal={arXiv preprint arXiv:2406.11794}
}

When using DCLM evaluation suite, please make sure to cite all the original evaluation papers. evaluation_bibtex.

When using DCLM for training, please make sure to cite the main training framework dependencies as well. training_bibtex.

License

This project is licensed under the MIT License. See the license file for details.