← Lessons

quiz vs the machine

Gold1410

Machine Learning

The Experiment Tracking Discipline

Log every run so past results stay comparable and recoverable.

5 min read · core · beat Gold to climb

Memory does not scale

After fifty experiments you will not recall which config produced the best score. Experiment tracking records the inputs and outputs of every run so they can be compared and reproduced.

  • Log the config: hyperparameters, data version, code commit.
  • Log the outputs: metrics, curves, and artifacts.
  • Tie each run to a unique id you can find again.

What good tracking buys

A clean log turns a pile of runs into a searchable history. You can sort by metric, diff two configs, and rebuild any model.

  • Compare runs fairly because the context is recorded.
  • Reproduce a result months later from its logged commit and data.
  • Avoid repeating experiments you already ran.

The record

The log is the project's institutional memory.

Key idea

Experiment tracking records each run's config, code version, data version, and metrics so results stay comparable and reproducible long after memory fades.

Check yourself

Answer to earn rating on the learn ladder.

1. What should experiment tracking record for each run?

2. Why is tracking the code commit and data version important?