Reading the Loss Curves

Your best guide is the pair of curves: training loss and validation loss over steps. Read them together. They tell you which failure mode you are in — and when to stop.

Intuition

Pattern	What it means	What to try
Underfitting	Both losses stay high and plateau early	Train longer, improve data, or give the model more room to learn
Overfitting	Train loss keeps falling while validation loss turns back up	Stop earlier, simplify updates (freeze / smaller learning rate), add better or more diverse data
No learning	Both curves flat near the starting value	Check data pipeline, labels, learning rate, and that gradients are actually flowing
Healthy fine-tune	Both losses fall together with a small, stable gap	Stop near the validation-loss minimum

Here is what those four patterns actually look like as numbers:

Overfitting                      Healthy
step  train   val                step  train   val
100   1.80   1.85                100   1.80   1.86
300   1.20   1.35                300   1.21   1.30
500   0.70   1.28                500   0.88   1.05
700   0.35   1.41  <- val rose   700   0.71   0.94
900   0.12   1.66  <- worse      900   0.66   0.91  <- flattening

On the left, everything after step 500 made the model worse for real users while looking better on the training chart. The checkpoint worth keeping is step 500.

How it works

What good learning looks like

Hallmarks of a healthy run:

Curves move together — Validation follows training downward; the model is generalizing, not only memorizing.
Small, stable gap — A modest train–val gap that stops widening is normal.
Diminishing returns — Loss flattens smoothly; you do not see wild oscillation, spikes, or divergence.
Stop at the val minimum — Checkpoint there. Training longer often buys nothing (or buys overfitting).

Tie-back to earlier chapters

Spikes and divergence → revisit the training instability checklist.
Task up, general skills down → revisit catastrophic forgetting canaries (loss curves alone will not show that).
Curves look sick despite a sane schedule → revisit data preparation.

When the curve looks fine but something is wrong

Two situations deserve suspicion even when the chart is pretty:

Validation loss falls unusually fast and low. Check whether validation examples leaked into the training file. Duplicated rows across the two sets make the model look brilliant and tell you nothing.
Loss looks great, outputs look worse. Loss measures next-token probability, not usefulness. A model can get better at predicting your training phrasing while becoming more repetitive or less helpful. Read a handful of real generated outputs at each checkpoint, not just the numbers.

What goes wrong

Shipping the checkpoint with the lowest training loss.
Ignoring validation until the end of a long run.
Reading only one curve and calling it “done.”

One-line summary

Watch training and validation loss together: fall together with a small gap means healthy learning; stop at the validation minimum.

Key terms

Underfitting — Model never gets good on train or validation.
Overfitting — Train looks great; validation gets worse.
Early stopping — Halt when validation stops improving.
Validation loss — Loss on held-out data used to pick checkpoints.