When you train a model, two numbers tell you most of the story: training loss and validation loss, recorded once per epoch. Plotted together as a chart, they show whether the model is still learning, has stopped improving, or has started to memorize the training data. This guide shows how to read that chart, and how to make one without writing any code.
What the two lines mean
Training loss measures how well the model fits the data it learns from. Validation loss measures how well it does on data it has never seen, which is the number that matters in the real world. In a healthy run, both lines fall together and then flatten out. If training loss keeps falling while validation loss starts rising, the model is overfitting: it is memorizing the training examples instead of learning patterns that generalize. If both lines stay high and flat, the model is underfitting: it is too simple, or it has not trained long enough, to capture the pattern.
How to plot training and validation loss without code
- Collect your numbers. Write down the training loss and validation loss for each epoch, in epoch order, starting at epoch 1.
- Paste each series as comma-separated values. Keep the two lists the same length, one value per epoch, with no missing epochs in the middle.
- Draw both lines on one chart. Put Epoch on the bottom axis and Loss on the side axis, with one line for training loss and one for validation loss.
- Read the shape, not single points. One wobbly epoch means little. The trend across many epochs is what tells you whether learning is working.
- Mark the turning point. If validation loss bottoms out and then rises while training loss keeps falling, note that epoch: it is the point where more training started to hurt.
A worked example
Suppose your training losses over five epochs are 1.2, 0.8, 0.6, 0.5, 0.45, and your validation losses are 1.3, 0.9, 0.7, 0.75, 0.9. Plotted together, both lines fall for three epochs, then the validation line turns upward while the training line keeps falling. That is the classic overfitting shape. The sensible next step is to stop around epoch 3, or to add regularization, gather more data, or simplify the model, rather than training longer and hoping the validation line recovers.
Mistakes to avoid
- Do not judge a run from the final epoch alone. A low training loss with a high validation loss is a warning, not a success.
- Do not mix up the two series. Swapping them flips the story and can hide overfitting completely.
- Do not leave gaps in the lists. A missing epoch shifts every later point and bends the chart.
- Do not smooth the lines so much that the turning point disappears. The moment validation loss turns upward is the most useful thing on the chart.
Try it faster with the free tool
The free ML Training Progress Plotter below draws the chart for you: paste each series as comma-separated values, in epoch order, click one button, and both lines appear together, with Epoch on the bottom axis and Loss on the side axis. Use it to see if a model is still learning or starting to overfit.
Try it here: ML Training Progress Plotter.
Two lines, one chart, and one question: is validation loss still falling? If it is not, more training is not the answer.
Questions readers ask
How many epochs do I need? More is better, up to a point. With only three epochs, a trend is hard to see. With twenty or more, the shape of the two lines becomes clear, and the turning point is easy to spot.
What if the lines are noisy? Small ups and downs from epoch to epoch are normal. Judge the overall direction, not one spike. If the noise hides the trend, record the loss averaged over a few epochs instead.
Does a low loss always mean a good model? No. A model can reach a very low training loss while failing on new data. That is why validation loss is on the chart: it is the honest check.