What Is a Training Curve? Reading a Model's Progress at a Glance
A plot of loss over time is the most basic diagnostic tool in training — its shape reveals whether a run is progressing well long before training finishes.
Loss, plotted over the course of training
A training curve is a plot of a model's loss — usually both training loss and validation loss — over the course of a training run, giving a visual, at-a-glance summary of how training is progressing at any given point.
What a healthy curve looks like
A healthy training curve shows loss decreasing steadily and smoothly, with training and validation loss tracking each other reasonably closely — a widening gap between the two, a sudden jump, or a curve that flattens out prematurely are all recognizable warning signs visible directly in the plot's shape.
Why teams watch it in real time, not just after the fact
Because a frontier training run can take weeks, catching a problem early — a loss spike, a plateau, a diverging gap between training and validation loss — by monitoring the curve live can save enormous amounts of wasted compute compared to only checking after the run finishes.
Frequently Asked Questions
What is a training curve?
A plot of a model's loss — typically both training and validation loss — over the course of a training run, giving an at-a-glance visual summary of how training is progressing.
Why do teams monitor training curves in real time?
Because a frontier training run can take weeks, catching a problem early — a loss spike, a plateau, a diverging training-validation gap — by watching the curve live can save enormous amounts of wasted compute versus checking only afterward.