Skip to main content

Exploring results

Find runs and inspect their settings, measurements, and files.

Use the Runs page to find a result and understand how it was produced. Each run records a piece of work, such as training a model, measuring its accuracy, or profiling it on a device. Open a run to inspect the settings, measurements, and files its author recorded.

Find a run#

Open your project’s Runs page.

  1. Use search or Filter to narrow the list by name, run type, or tags.
  2. Open Columns to show the metrics, parameters, and tags you need.
  3. Select a run’s name to open it.

Use names that describe the work, such as INT8 evaluation on validation_v1, and tags to group related candidates. To search across projects you can access, select All projects and open Runs.

Inspect a result#

Read the measurement alongside the settings that produced it.

This evaluation records 0.92 accuracy, or 92%. Its parameters identify the INT8 model and validation_v1 dataset, so you can tell which candidate was tested and what the score describes.

Artifacts hold supporting files such as the model, an evaluation report, or plots. Tags help organize runs; Links connect them to work in other tools. Use Dashboard for charts attached to the run and Lineage to follow the work that produced it.

Record your reasoning#

Use the run’s Description to record what you tried, why you chose the settings, and what you learned. Include alternatives you ruled out and links to related runs or reports. These notes help teammates understand the work and give you context to draw on when submitting evidence.

Compare alternatives#

Open Experiments, select the runs, and choose a dashboard. Use a table for exact values, a line chart for progress over steps, or a scatter plot for a tradeoff such as accuracy and latency. See Dashboards.

Select a baseline run alongside the candidates you want to assess. Check which conditions changed: model, dataset, device, runtime, or batch size. Keep the other conditions consistent so you can understand the effect of that change.

Compare the same measurements and units. For example, p95 latency and mean latency are different measurements, even when both use milliseconds. If your evaluation records results for individual classes or conditions, inspect those alongside the overall score. An average can hide a regression in a particular condition.

Next steps#