Nice to see a practical walkthrough of pretraining instead of only fine-tuning examples. Most people don’t realize how much of the work sits in data prep, stable training loops, and managing compute. Even a small BERT run is a good reminder of the gap between theory and an actual training pipeline.