Model Configuration

Configuring a model source can be done by specifying a mapping with a python object. It allows you to point to specific models you have created or gives the ability to access hugging face models.
Specify a Model discusses how to create a Python model file.

For an example of how to use this config in the model registry, please see the “Register a Model” step in Stress Tests walkthrough.

Template

Model Path

{
  "model_path": {
    "path": "path/to/model.py"
  }
}

Huggingface

{
  "hugging_face" : {
    "model_uri": "cross-encoder/nli-MiniLM2-L6-H768",
    "kwargs": "{\"tokenizer_uri\": \"cross-encoder/nli-MiniLM2-L6-H768\", \"class_map\": {\"0\": \"contradiction\", \"1\": \"entailment\", \"2\": \"neutral\"}, \"ignore_class_names\": true}"
  }
}

Arguments

Parameter	Type	Description
model_path	JSON object	Specification for a model provided via a model file.
path	String	Path to the Python model file.
hugging_face	JSON object	Specification for a huggingface model.
model_uri	String	The URI to a HuggingFace model.
kwargs	String	A JSON String of arguments to pass to HuggingFace model.

Specify exactly one of model_path or hugging_face.