Fine-tuned model performance degradation after 3 epochs

I'm fine-tuning GPT-4o-mini on a customer service dataset (5000 examples) and seeing performance peak at epoch 2-3, then degrade significantly.

Training metrics

• Epoch 1: Loss 1.23, eval accuracy 78%
• Epoch 2: Loss 0.87, eval accuracy 84%
• Epoch 3: Loss 0.62, eval accuracy 86%
• Epoch 4: Loss 0.41, eval accuracy 79% (degrading!)
• Epoch 5: Loss 0.28, eval accuracy 72%

Classic overfitting curve. My questions: 1. Is there a way to set early stopping in the fine-tuning API? 2. Should I increase my dataset size, or is 5000 sufficient? 3. Any recommendations for learning rate multiplier?