Training jobs: metadata.yaml

The metadata.yaml file defines how Modelbit should run your training job. It specifies the Python version, argument types and schedule behavior.

File format

The file is YAML-formatted. Here are the type definitions and field descriptions:

training_jobs/example_job/metadata.yaml

owner: str
runtimeInfo: Dict[str, Any]
  mainFunction: str
  mainFunctionArgs: List[str]
  pythonVersion: Literal["3.6" | "3.7" | "3.8" | "3.9" | "3.10" | "3.11" | "3.12"]
  baseImage: Optional[str]
  systemPackages: Optional[List[str]]
schedules: Optional[List[Dict[str, Any]]]
  - beforeStart: Optional[Dict[str, Any]]
      refreshDatasets: List[str]
    onFailure: Optional[Dict[str, Any]]
      sendEmail: str
    schedule: cron-string
    size: Literal["small", "medium", "large", "xlarge", "2xlarge", "4xlarge", "gpu_small", "gpu_medium", "gpu_large"]
schemaVersion: Literal[1]

The fields:

• owner: The email you're using in Modelbit. The job will run as the owner's role. If the user no longer exists in Modelbit the job will stop running.

• runtimeInfo:

  • mainFunction: The name of the function to call in source.py with the arguments sent to your training job.
  • mainFunctionArgs: The argument names and types used by mainFunction, formatted as <arg-name>:<arg-type>. There should not be spaces in these values (i.e. use foo:int, not foo: int).
  • pythonVersion: The major-minor version of Python to use.
  • baseImage: The name of the base image to use in this training job's environment. Learn more about base images.
  • systemPackages: A list of packages to install in the production environment with apt-get.

• schedules: Optional list of schedule dicts. Each schedule can have the following fields:

  • beforeStart:
    • refreshDatasets: List of dataset names to refresh before beginning the training job.
  • onFailure:
    • sendEmail: Email address to alert if the training job exits unsuccessfully.
  • schedule: Cron string representing the schedule of the training job.
  • size: Size of the machine to use when running the job.

• schemaVersion: The file format version of this file.

Examples

Here are example metadata.yaml files showing various popular configurations.

Minimum metadata.yaml

The main function in the training job:

source.py

def train_my_predictor():
    return ...

The associated metadata.yaml:

metadata.yaml

owner: you@company.com
runtimeInfo:
  mainFunction: train_my_predictor
  mainFunctionArgs: []
  pythonVersion: "3.9"
schemaVersion: 1

Specifying system packages to install

Use the systemPackages key within runtimeInfo to specify which packages to install with apt-get:

source.py

def train_my_predictor(a: int) -> int:
    return ...

The associated metadata.yaml:

metadata.yaml

owner: you@company.com
runtimeInfo:
  mainFunction: train_my_predictor
  mainFunctionArgs:
    - a:int
    - return:int
  pythonVersion: "3.9"
  systemPackages:
    - build-essential
    - libgl1
    - libgl1-mesa-glx
    - libglib2.0-0
schemaVersion: 1

Daily schedule with dataset refresh

This schedule runs every midnight UTC (0 0 * * *) and refreshes the dataset my_dataset before beginning the training job.

metadata.yaml

owner: you@company.com
runtimeInfo:
  mainFunction: train_my_predictor
  mainFunctionArgs: []
  pythonVersion: "3.10"
  systemPackages: null
schedules:
  - beforeStart:
      refreshDatasets:
        - my_dataset
    onFailure:
      sendEmail: you@company.com
    schedule: 0 0 * * *
    size: medium
schemaVersion: 1

Daily schedule passing arguments

This training job takes two arguments (a and b). The scheduled execution of this job passes 25 and 42 as arguments to train_my_predictor.

source.py

def train_my_predictor(a: int, b: int) -> int:
    return ...

The associated metadata.yaml:

metadata.yaml

owner: you@company.com
runtimeInfo:
  mainFunction: train_my_predictor
  mainFunctionArgs:
    - a:int
    - b:int
  pythonVersion: "3.10"
  systemPackages: null
schedules:
  - arguments:
      - 25
      - 42
    schedule: 0 0 * * *
    size: large
schemaVersion: 1

Two schedules

This training job has two schedules. One nightly (passing the argument 1) and one weekly (passing the argument 2):

metadata.yaml

owner: you@company.com
runtimeInfo:
  mainFunction: train_my_predictor
  mainFunctionArgs:
    - a:int
  pythonVersion: "3.10"
  systemPackages: null
schedules:
  - arguments:
      - 1
    schedule: 0 0 1 * *
    size: small
  - arguments:
      - 2
    schedule: 0 0 * * SUN
    size: small
schemaVersion: 1