FairPy: A Toolkit for Evaluation of Social Biases and their Mitigation in Large Language Models (2302.05508v1)

Abstract: Studies have shown that large pretrained LLMs exhibit biases against social groups based on race, gender etc, which they inherit from the datasets they are trained on. Various researchers have proposed mathematical tools for quantifying and identifying these biases. There have been methods proposed to mitigate such biases. In this paper, we present a comprehensive quantitative evaluation of different kinds of biases such as race, gender, ethnicity, age etc. exhibited by popular pretrained LLMs such as BERT, GPT-2 etc. and also present a toolkit that provides plug-and-play interfaces to connect mathematical tools to identify biases with large pretrained LLMs such as BERT, GPT-2 etc. and also present users with the opportunity to test custom models against these metrics. The toolkit also allows users to debias existing and custom models using the debiasing techniques proposed so far. The toolkit is available at https://github.com/HrishikeshVish/Fairpy.