Loss is how badly the model matches the expected outputs, and it should fall smoothly. The shape of the fall tells you whether the run is healthy, and three shapes cover almost every problem a tuning run has.
About 1 min of reading
Spot the shapes.
Healthy
Before you run it, say out loud what it will print. Then open the answer.
Contradiction
Before you run it, say out loud what it will print. Then open the answer.
Read the first example one line at a time. A line that prints shows what appears on screen; the rest set things up quietly.
Training loss falling while held out loss rises is overfitting. That pair of lines is the single most useful plot, and it is the one that tells you to stop earlier.
A spiky curve is usually contradictory data. Finding the contradiction in the set is the fix; lowering the rate hides it and leaves a model that still gets those cases wrong.
A steady decrease is a healthy run. It says the signal in the set is consistent, which is the thing you want to know before you serve the model.
A loss that does not move means the model is not learning: the set is too small, the rate is too low, or the format is wrong. A flat curve is a bug in the setup, not a bad model.
A jagged curve means the updates disagree with each other, usually from a set with contradictory examples. Two examples that want different outputs for the same input produce exactly this.
Each one is defined here on this site. Open the entry for the full definition and the other words it points at.