Epoch
One full pass of the training set through the model. Fine-tuning runs are usually 1-4 epochs; each extra epoch multiplies the GPU cost of the run and, past a point, causes overfitting.
What it is
If your dataset has 10,000 examples and you train for 3 epochs, the model sees 30,000 examples total (in a different random order each time). Wall-clock time and GPU cost scale almost linearly with epoch count.
Why it matters
Epoch count is the dial with the most direct effect on run cost. More is not better: 1-3 epochs is standard for LoRA fine-tuning, and validation loss usually starts rising after that.