Implementation:NVIDIA NeMo Curator ModelInterface

Knowledge Sources	NVIDIA NeMo Curator
Domains	Machine Learning, Architecture, Interfaces
Last Updated	2026-02-14 00:00 GMT

Overview

Abstract base class that defines the interface contract for all machine learning models in the NeMo Curator pipeline, focused on weight handling and environmental setup.

Description

The ModelInterface class uses Python's abc.ABC to declare two abstract members that all model wrappers must implement:

model_id_names -- An abstract property that returns a list of model identifier strings. These are typically HuggingFace-style model IDs (e.g., Salesforce/instructblip-vicuna-13b) that uniquely identify the models associated with the implementation.
setup() -- An abstract method for loading weights and building computation graphs. This is called after construction to prepare the model for inference.

The interface is intentionally minimal. It standardizes weight management and environment setup without constraining how inference is performed. This allows each concrete model implementation to define its own __call__, generate, or other inference methods as appropriate.

Usage

Implement ModelInterface when creating a new model wrapper for the NeMo Curator pipeline. The pipeline infrastructure relies on model_id_names for weight downloading and setup() for model initialization, enabling uniform handling of all model types across the system.

Code Reference

Source Location

Repository: NeMo-Curator
File: nemo_curator/models/base.py
Lines: 1-42

Signature

class ModelInterface(abc.ABC):
    @property
    @abc.abstractmethod
    def model_id_names(self) -> list[str]:
        """Returns a list of model IDs associated with the model."""

    @abc.abstractmethod
    def setup(self) -> None:
        """Set up the model for use, such as loading weights and building computation graphs."""

Import

from nemo_curator.models.base import ModelInterface

I/O Contract

Abstract Members

Name	Type	Description
model_id_names	property -> list[str]	Returns a list of HuggingFace-style model identifiers associated with this model implementation
setup()	method -> None	Initializes the model by loading weights and building computation graphs; must be called before inference

Known Implementations

The following classes in the NeMo Curator codebase implement ModelInterface:

Class	Module	Purpose
AestheticScorer	nemo_curator.models.aesthetics	Aesthetic quality scoring from CLIP embeddings
NSFWScorer	nemo_curator.models.nsfw	NSFW content detection from CLIP embeddings
QwenLM	nemo_curator.models.qwen_lm	Text-only language model for caption enhancement
QwenVL	nemo_curator.models.qwen_vl	Vision-language model for video captioning

Usage Examples

Basic Usage

from nemo_curator.models.base import ModelInterface

class MyCustomModel(ModelInterface):
    def __init__(self, model_dir: str) -> None:
        self.model_dir = model_dir
        self.model = None

    @property
    def model_id_names(self) -> list[str]:
        return ["my-org/my-model-name"]

    def setup(self) -> None:
        # Load weights and prepare the model
        self.model = load_model(self.model_dir)
        self.model.eval()

    def __call__(self, inputs):
        return self.model(inputs)

Related Pages

Environment:NVIDIA_NeMo_Curator_Python_Linux_Base
NVIDIA_NeMo_Curator_AestheticScorer -- Concrete implementation for aesthetic scoring
NVIDIA_NeMo_Curator_NSFWScorer -- Concrete implementation for NSFW scoring
NVIDIA_NeMo_Curator_QwenLM -- Concrete implementation for Qwen text LM
NVIDIA_NeMo_Curator_QwenVL -- Concrete implementation for Qwen vision-language model

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment