Data Quality

You have 100K noisy training examples or 5K curated ones. Research shows the smaller clean set usually wins. What principle does this demonstrate? Flag format: CONGRESS{[principle]} Example: CONGRESS{speed_over_accuracy}