Comment by energy123

80% to 100% would be an even smaller improvement but arguably the most impressive and useful (assuming the benchmark isn't in the training data)