Could it be that text sources are plenty, and more dense than training for videos, and images?