Experiment Tracking
Recording every training run's code version, data, hyperparameters, metrics and resulting model, so results can be compared, reproduced and safely deployed.
Also MLflow · Weights & Biases · model registry · reproducibility · run tracking
A lab notebook for machine learning. Every time you try a recipe, you write down exactly what you used and how it turned out, so you can repeat the best one later.
Where it fits
+All relationships as a list
How it works
Machine learning is experimental. You change the learning rate, add a feature, try a new dataset version, and after a week you have dozens of runs. Which one was the good one, and what exactly went into it?
Experiment tracking answers that by logging, for every run:
- Code: the git commit.
- Data: which version of the dataset.
- Configuration: all hyperparameters.
- Metrics: training and validation curves, final scores.
- Artifacts: the trained model, plots, sample predictions.
Dashboards then let you sort and compare runs. A model registry adds a controlled path from “best experiment” to “approved for serving”, with versions you can roll back.
It sounds like bookkeeping, and it is, but it is the difference between “we think the model got better” and being able to prove why. Regulations such as the EU AI Act also require documentation for high-risk systems.
Everyday use cases
- Case 01
Hyperparameter searches
Comparing 200 runs side by side to see which learning rate and model size worked best.
- Case 02
Audits
Answering "which data and code produced the model that made this decision?" months later.
Strengths, weaknesses and when to use it
Strengths
- Cheap insurance against lost results and unreproducible models.
Weaknesses
- Only useful if logging is consistent; adds a small amount of setup.
When to use itFrom the very first experiment of any project that may last more than a day.
Compare with
Check yourself
- Q01
Which detail is most important to log to reproduce a run?
Show answer
A: Code version, data version and hyperparameters