Do small language models know what they don't know?