Appendix Uncovering and Quantifying Social Biases in Code Generation

We conduct a preliminary study on finding a proper prompt construction strategy. In this section, we quantify the efficacy of different code prompts to elicit social biases in pre-trained code generation models. We mainly study the following aspects: the number of functions contained in the prompt, the relevancy of functions to humans, and the order of functions in the code prompt. Experimental results are shown in Table 1. As we can see in the table, CBS increases with the number of functions both for InCoder and CodeGen.