Honestly, it was a budget I set up front. The command was --iters 300 with an eval and a checkpoint saved every 100 steps. I only picked the step-100 checkpoint afterwards, once I saw val loss go 1.382 → 1.665 → 1.811 while train loss kept dropping to 0.046.

Looking back, 300 was far too many for the data. It's 108 training examples at batch size 4, so about 27 steps per epoch. 300 steps is roughly 11 passes over the same data, and step 100 is already about 3.7. The validation set is also small (only 12 examples), and I only checked it every 100 steps. So 100 is just the best of the three points I measured, not the actual lowest point. The real minimum could be anywhere from about step 50 to 150.

Next time I'll run eval every 25 steps and stop when val loss stops improving. Thanks for raising it. It's an easy one to miss when the training curve looks so good.