Check your checkpoints/ folder right now. Do you have a V3? Or did you let the auto-saver overwrite it with V18 garbage?
We are afraid of overwriting a happy accident. We are afraid that once we save, we have to admit the work is real. But here is the truth: V2 is already dead. save v3
You just made a deal with a witch. You sold your best sword. The dragon is angry. Check your checkpoints/ folder right now
Epoch 2: The model overfits the noise. It is confident, but wrong. You save it out of hope. We are afraid of overwriting a happy accident
Never overwrite V3 in a panic.
In machine learning, V3 (or epoch 3 with the right learning rate) is often where the "Ugly Duckling" turns into a swan. The loss curve bends. The validation accuracy spikes.