Characteristics of Harmful Text: Towards Rigorous Benchmarking of Language Models