The gesture
- Train a model.
- Open Would more data help? below the leaderboard.
- Choose a model and trace the curve.
LabML retrains that model on growing, nested fractions of the training set, and scores each one on the same full test set.
How to read it
Look at the right edge of the curve.
| What you see | What it means | What to do |
|---|---|---|
| It is still climbing | The model has not finished learning from your data | Collecting more rows is worth it |
| It is flat | More rows will change nothing | Work on the features, or change model |
| It climbs then falls | Rare; often a sign of leakage or a doubtful split | Check the leak detector |
The band around the curve is a bootstrap interval. If it is wide at the right edge, the slope you think you see may be noise.
Why nested fractions
Each size is a prefix of the next: the first 200 rows are inside the first 400. Without that, each point would draw a different sample and the curve would measure the luck of the draw as much as the effect of size.
What the curve does not say
It does not say how many rows would be needed. Extrapolating a learning curve beyond what was measured is a guess, and LabML does not offer one.
Where to go next
- Compare two runs to measure a change of features.
- The method choices on why the baseline comes first.