Researchers release dataset to expose racial, religious, and gender biases in language models