What are "model artifacts" in MLflow?

Model artifacts in MLflow refer to the files created during the training process that capture the necessary elements of a machine learning model. These artifacts can include various components such as the trained model itself (the weights and configurations used), any pre-processing steps, or configurations that define how the model was built.

When you train a model, a number of files are generated that represent the model's readiness to be deployed or served. These could comprise serialized model files in formats like PMML, ONNX, or native formats depending on the framework used. Additionally, model artifacts may also include any metadata related to the training run, such as training parameters, metrics, or even the Python code that was used to develop the model.

This understanding is crucial because model artifacts are central to ensuring reproducibility in machine learning experiments. They allow other data scientists or systems to easily use the trained model later on without needing to retrain it from scratch. In contrast, the other options do not accurately capture the essence of model artifacts. For instance, textual documentation, predictions, or visualization graphs do not specifically represent the physical or digital files generated during the training process.