If you have 10,000 examples, one epoch is when the model has learned from each of them once. Fine-tuning usually needs 1–3 epochs.
More epochs mean more time and cost; too many and the model memorises instead of learning (overfitting).
Example
2 million tokens of data × 3 epochs = 6 million tokens of training.
বাংলায়: ইপক (epoch) — পুরো ট্রেনিং ডেটা মডেলকে একবার দেখানো = এক ইপক।