An epoch is one complete pass through every entry in a training dataset. Fine-tuning commonly repeats this pass several times so the model can progressively adjust toward the desired responses. section 12.2.2
During a fine-tuning run, the epoch value indicates how far the process has progressed through the current pass. It is therefore a progress measure, not a quality score by itself. section 12.2.4
Reading epoch progress
A training log such as {"loss": 1.0936, "learning_rate": 0.00019999, "epoch": 0.02} shows that the run is only 0.02 epochs into its dataset pass. Repeated log entries at the same displayed value can report different loss values while progress remains near the beginning. section 12.2.4
{'loss': 1.0936, 'learning_rate': 0.00019999154711147226, 'epoch': 0.02}
{'loss': 1.3172, 'learning_rate': 0.00019999135609452385, 'epoch': 0.02}
{'loss': 1.0351, 'learning_rate': 0.0001999911629434316, 'epoch': 0.02}
Inspect representative training-log entries showing epoch progress alongside loss and learning rate. section 12.2.4
