Scaling Laws For Neural Language Models