Model Registry

A Model Registry is a centralized catalog that stores and versions trained models along with their metadata — training dataset hash, hyperparameters, evaluation metrics, environment dependencies, and author — enabling reproducibility, governance, and deployment automation. Each registered model has: a unique name and version (semantic or auto-increment), a stage tag (Staging, Production, Archived), stored artifacts (model weights, tokenizer, config), and lineage links (dataset version, training run ID, git commit hash). MLflow Model Registry is the most widely adopted open-source implementation; commercial alternatives include AWS SageMaker Model Registry and Vertex AI Model Registry.

A model registry is a central catalog of all models — like a package registry (npm, PyPI) but for trained models. It stores model artifacts, metadata (training data, hyperparameters, metrics), and deployment status (staging, production, archived). MLflow Model Registry is the most common.

Model registry usage: (1) register model: mlflow.register_model("runs:/<run_id>/model", "my-model"), (2) stages: Staging (testing), Production (live), Archived (deprecated), (3) deployment fetches model by stage: "give me the production version of my-model," (4) audit: every model has a history — who trained it, when, with what data, what metrics, (5) for LLMs: registry tracks base model + adapter (LoRA) combinations, quantization levels, and GGUF versions.