Skip to main content

Model Registry

Group training, evaluation, and profiling runs by model.

Use Model Registry to organize the runs for each model candidate. Teammates can find how it was produced and evaluated. Each run keeps its own settings, measurements, and artifacts, including model files.

This INT8 candidate reaches 92% accuracy and 18.4 ms p95 latency. The two measurements come from separate evaluation and profiling runs.

Read the model’s results#

The Overview brings together parameters, metrics, and artifacts from the associated runs. Here, accuracy and latency help you assess whether the quantized classifier is a suitable candidate for the target device.

Open a source run to see the dataset, device, and settings behind a measurement. Use Dashboard to explore the results in charts, or Runs to browse the individual evaluations. Lineage shows how the recorded work is connected.

Create a model#

Open Models in your project and select Create model. Choose the runs that evaluate the same candidate:

The evaluation and profile runs describe different properties of the same model.

Continue through the lineage preview, then give the model a name and description. Describe its intended use so teammates know what they are assessing. You can add or change the associated runs later through Edit in the model actions.

Make a candidate easy to find#

Pin a model to the project home page when it is a useful reference for the team, such as the current baseline or a candidate ready for review. Share its URL to point teammates to the results behind it.

A model gives those results a named home. Use a run group for a collection of executions, and lineage to follow how one run led to another.

Next steps#