Synthetic data is cheap for fine-tuning, but too much of it degrades the model. Where is the tipping point in your experience?