The gesture
- Train once. The run is saved to the history at the bottom of the ML Lab.
- Change one single thing — remove a column, change the split, enable class weighting.
- Train again.
- In the history, select both runs and open Compare.
Up to six runs can be compared at once.
What the page gives you
- The configuration on both sides, with the differences highlighted.
- The features added and removed, named.
- The per-model leaderboard: each family, its score on both sides, and the gap.
- Cross-run uncertainty: does the gap survive resampling, or does it sit inside the noise?
That last line is the one that counts. A +0.01 gap over 178 test rows is not a gain — it is a draw.
Why change one thing only
Because the seed is fixed, everything left unchanged stays identical. If you change one variable, the gap is attributable to it.
If you change three, the page will faithfully show you a gap you will not be able to apportion. LabML cannot untangle that for you.
Where to go next
- Read a learning curve if the question is « would more data help? ».
- The method choices on why uncertainty is shown rather than hidden.