Skip to content

CMF in a nutshell

CMF (Common Metadata Framework) collects and stores information associated with Machine Learning (ML) pipelines. It also implements APIs to query this metadata. The CMF adopts a data-first approach: all artifacts (such as datasets, ML models and performance metrics) recorded by the framework are versioned and identified by their content hash.


1. Pre-Requisites:

  • 3.9>= Python <=3.10
  • Git latest version

2. Set up Python Virtual Environment:

conda create -n cmf python=3.10
conda activate cmf
virtualenv --python=3.10 .cmf
source .cmf/bin/activate

3. Install CMF:

pip install git+
# pip install cmflib

Next Steps

After installing CMF, proceed to configure CMF server and client. For detailed configuration instructions, refer to the Quick start with cmf-client page.


Complex ML projects rely on ML pipelines to train and test ML models. An ML pipeline is a sequence of stages where each stage performs a particular task, such as data loading, pre-processing, ML model training and testing stages. Each stage can have multiple Executions. Each Execution,

  • consume inputs and produce outputs.
  • are parametrized by parameters that guide the process of producing outputs.

ML Pipeline Definition Example

CMF uses the abstractions of Pipeline,Context and Executions to store the metadata of complex ML pipelines. Each pipeline has a name. Users provide it when they initialize the CMF. Each stage is represented by a Context object. Metadata associated with each run of a stage is captured in the Execution object. Inputs and outputs of Executions can be logged as dataset, model or metrics. While parameters of executions are recorded as properties of executions.

CMF abstractions

Start tracking the pipeline metadata by initializing the CMF runtime. The metadata will be associated with the pipeline named test_pipeline.

from cmflib.cmf import Cmf
from ml_metadata.proto import metadata_store_pb2 as mlpb

cmf = Cmf(

Before we can start tracking metadata, we need to let CMF know about stage type. This is not yet associated with this particular execution.

context: mlmd.proto.Context = cmf.create_context(

Now we can create a new stage execution associated with the train stage. The CMF always creates a new execution, and will adjust its name, so it's unique. This is also the place where we can log execution parameters like seed, hyper-parameters etc .

execution: mlmd.proto.Execution = cmf.create_execution(
    custom_properties = {"num_epochs": 100, "learning_rate": 0.01}

Finally, we can log an input (train dataset), and once trained, an output (ML model) artifacts.

    'artifacts/test_dataset.csv',   # Dataset path 
    "input"                         # This is INPUT artifact
    "artifacts/model.pkl",          # Model path 
    event="output"                  # This is OUTPUT artifact

ML Pipeline Stage Execution

Quick Example

Go through Getting Started page to learn more about CMF API usage.

API Overview

Import CMF.

from cmflib import cmf

Initialize CMF. The [CMF][cmflibcmfcmf] object is responsible for managing a CMF backend to record the pipeline metadata. Internally, it creates a pipeline abstraction that groups individual stages and their executions. All stages, their executions and produced artifacts will be associated with a pipeline with the given name.

cmf = cmf.Cmf(
   filename="mlmd",                # Path to ML Metadata file.
   pipeline_name="mnist"           # Name of a ML pipeline.

Define a stage. An ML pipeline can have multiple stages, and each stage can be associated with multiple executions. A stage is like a class in the world of object-oriented programming languages. A context (stage description) defines what this stage looks like (name and optional properties), and is created with the create_context method.

context = cmf.create_context(
    pipeline_stage="download",     # Stage name
    custom_properties={            # Optional properties
        "uses_network": True,      #  Downloads from the Internet
        "disk_space": "10GB"       #  Needs this much space

Create a stage execution. A stage in ML pipeline can have multiple executions. Every run is marked as an execution. This API helps to track the metadata associated with the execution, like stage parameters (e.g., number of epochs and learning rate for train stages). The stage execution name does not need to be the same as the name of its context. Moreover, the CMF will adjust this name to ensure every execution has a unique name. The CMF will internally associate this execution with the context created previously. Stage executions are created by calling the create_execution method.

execution = cmf.create_execution(
    execution_type="download",            # Execution name.
    custom_properties = {                 # Execution parameters
        "url": ""   #  Data URL.

Log artifacts. A stage execution can consume (inputs) and produce (outputs) multiple artifacts (datasets, models and performance metrics). The path of these artifacts must be relative to the project (repository) root path. Artifacts might have optional metadata associated with them. These metadata could include feature statistics for ML datasets, or useful parameters for ML models (such as, for instance, number of trees in a random forest classifier).

  • Datasets are logged with the log_dataset method.

    cmf.log_dataset('data/mnist.gz', "input", custom_properties={"name": "mnist", "type": 'raw'})
    cmf.log_dataset('data/train.csv', "output", custom_properties={"name": "mnist", "type": "train_split"})
    cmf.log_dataset('data/test.csv', "output", custom_properties={"name": "mnist", "type": "test_split"})

  • ML models produced by training stages are logged using log_model API. ML models can be both input and output artifacts. The metadata associated with the artifact could be logged as an optional argument.

    # In train stage
       path="model/rf.pkl", event="output", model_framework="scikit-learn", model_type="RandomForestClassifier", 
    # In test stage
       path="model/rf.pkl", event="input" 

  • Metrics of every optimization step (one epoch of Stochastic Gradient Descent, or one boosting round in Gradient Boosting Trees) are logged using log_metric API.

    #Can be called at every epoch or every step in the training. This is logged to a parquet file and committed at the 
    # commit stage.
    #Inside training loop
    while True: 
         cmf.log_metric("training_metrics", {"loss": loss}) 

  • Stage metrics, or final metrics, are logged with the log_execution_metrics method. These are final metrics of a stage, such as final train or test accuracy.

    cmf.log_execution_metrics("metrics", {"avg_prec": avg_prec, "roc_auc": roc_auc})

Dataslices are intended to be used to track subsets of the data. For instance, this can be used to track and compare accuracies of ML models on these subsets to identify model bias. Data slices are created with the create_dataslice method.

dataslice = cmf.create_dataslice("slice-a")
for i in range(1, 20, 1):
    j = random.randrange(100)

Graph Layer Overview

CMF library has an optional graph layer which stores the relationships in a Neo4J graph database. To use the graph layer, the graph parameter in the library init call must be set to true (it is set to false by default). The library reads the configuration parameters of the graph database from cmf config generated by cmf init command.

cmf init minioS3 --url s3://dvc-art --endpoint-url http://x.x.x.x:9000 --access-key-id minioadmin --secret-key minioadmin --git-remote-url --cmf-server-url http://x.x.x.x:8080  --neo4j-user neo4j --neo4j-password password --neo4j-uri bolt://localhost:7687

Here, "dvc-art" is provided as an example bucket name. However, users can change it as needed, if the user chooses to change it, they will need to update the Dockerfile for minioS3 accordingly.

To use the graph layer, instantiate the CMF with graph=True parameter:

from cmflib import cmf

cmf =  cmf.Cmf(

Jupyter Lab docker container with CMF pre-installed

Use a Jupyterlab Docker environment with CMF pre-installed

CMF has a docker-compose file which creates two docker containers, - JupyterLab Notebook Environment with CMF pre installed. - Accessible at http://[HOST.IP.AD.DR]:8888 (default token: docker) - Within the Jupyterlab environment, a startup script switches context to $USER:$GROUP as specified in .env - example-get-started from this repo is bind mounted into /home/jovyan/example-get-started - Neo4j Docker container to store and access lineages.

Step 1.

create .env file in current folder using env-example as a template. Modify the .env file for the following variables USER,UID,GROUP,GID,GIT_USER_NAME,GIT_USER_EMAIL,GIT_REMOTE_URL #These are used by docker-compose.yml

Step 2.

Update docker-compose.yml as needed.

your .ssh folder is mounted inside the docker conatiner to enable you to push and pull code from git

Create these directories in your home folder

mkdir $HOME/workspace 
mkdir $HOME/dvc_remote 
workspace - workspace will be mounted inside the cmf pre-installed docker conatiner (can be your code directory)
dvc_remote - remote data store for dvc

Change the below lines in docker-compose to reflect the appropriate directories

 If your workspace is named "experiment" change the below line
$HOME/workspace:/home/jovyan/workspace to 
If your remote is /extmount/data change the line 
$HOME/dvc_remote:/home/jovyan/dvc_remote to 
Start the docker
docker-compose up --build -d
Access the jupyter notebook http://[HOST.IP.AD.DR]:8888 (default token: docker)

Click the terminal icon

Quick Start

cd example-get-started
cmf init local --path /home/user/local-storage --git-remote-url --cmf-server-url --neo4j-user neo4j --neo4j-password password --neo4j-uri bolt://localhost:7687
cmf artifact push
The above steps will run a pre coded example pipeline and the metadata is stored in a file named "mlmd".
The artifacts created will be pushed to configured dvc remote (default: /home/dvc_remote)
The stored metadata is displayed as image

Metadata lineage can be accessed in neo4j.
Open http://host:7475/browser/ Connect to server with default password neo4j123 (To change this modify .env file)

Run the query

MATCH (a:Execution)-[r]-(b) WHERE (b:Dataset or b:Model or b:Metrics) RETURN a,r, b     
Expected output

Jupyter Lab Notebook

Select the kernel as Python[conda env:python37]

Shutdown/remove (Remove volumes as well)

docker-compose down -v


CMF is an open source project hosted on GitHub and distributed according to the Apache 2.0 licence. We are welcome user contributions - send us a message on the Slack channel or open a GitHub issue or a pull request on GitHub.


    title={Self-Learning Data Foundation for Scientific AI},
    author={Martin Foltin, Annmary Justine, Sergey Serebryakov, Cong Xu, Aalap Tripathy, Suparna Bhattacharya, 
            Paolo Faraboschi},
    note = {Presented at the "Monterey Data Conference"},



Common Metadata Framework and its documentation are in active stage of development and are very new. If there is anything unclear, missing or there's a typo, please, open an issue or pull request on GitHub.