Training walks the set repeatedly. One pass is an epoch, and the number of passes is a knob you set. Too few leaves the model near its starting point, too many makes it repeat the training set, and the loss curve shows which is happening.
About 1 min of reading
Read a tiny curve.
The losses
Before you run it, say out loud what it will print. Then open the answer.
The stop
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.
Save the loss after each pass. Without the curve you have a number and no way to know whether the run stopped early or ran too long, which is the question the number alone cannot answer.
Evaluate on the held out set each pass and keep the checkpoint that scored best. That is the rule that turns 'how long do I train' into a measured answer.
A hundred examples seen once is a weak signal. The same hundred seen ten times is a strong one, which is why a small set usually needs more epochs than a large one.
Past the point where the set is learned, extra passes push the model toward copying those exact examples. That is overfitting, and it is visible as training loss falling while held out loss stops falling.
A larger batch is fewer updates from the same data. It changes speed and memory, not what is learned, so set it by what fits rather than by a guess about quality.
Each one is defined here on this site. Open the entry for the full definition and the other words it points at.