10.2 Experiment tracking
You can reproduce the run that produced your best number.
Experiment tracking records the parameters, metrics and artifacts of every run, so the best number is permanently attached to the exact configuration that produced it, with sweeps and comparison coming free. It sits early in the module because untracked experiments are unreproducible by definition. The selective-logging trap is the one to avoid: log the failures too, or in three weeks you will re-run them, because nothing recorded that they failed.
Work through these
MLflow or Weights & Biases setup
A service that records every run with its settings and results, so the best number can be traced back to what produced it. Two widely used options are covered.
MLflow Tracking Quickstart · TutorialLogging params, metrics, artifacts
What to record: the settings that defined the run, the numbers it produced, and the files it left behind. Recording too little is the common mistake and it is discovered too late.
Comparing runs and sweeps
Comparing runs side by side and searching over many at once is what turns a log into a tool. This is where the recording pays back.
Model registry and stage promotion
A registry holds the models considered good enough to use, with a marked stage for each. It is the boundary between experimentation and production.
Sign in to keep your progress.
Free resources
We haven't checked most of these for screen reader use yet.
Links last checked 29 Aug 2026.
Stuck here?
Ask a mentor. A real person answers, and they can see exactly which topic you're on. Usually within a couple of working days.
Checking your session…
Topics shown in module order.