FairLangProc: A Python package for fairness in NLP

The astonishing results of the transformer architecture on Natural Language Processing (NLP) tasks (Devlin et al. 2019; Radford et al. 2019), their scalation properties (Vaswani et al. 2017) and the massive amount of text data available (Wang et al. 2019; Foundation Accessed 27/05/2025) have led to the development of Large Language Models (LLM) whose performance towers above that of traditional Language Models (LM) (Zhang et al. 2021; BigScience et al. 2022). Furthermore, LLMs have been widely adopted for custom downstream tasks by leveraging the flexibility provided by fine-tuning (Chung et al. 2024) and their few-shot learning capabilities (Brown et al. 2020), establishing a new zeitgeist in the NLP community. These factors have led to their widespread adoption across major areas of society such as academia (Naveed et al. 2023; Meyer et al. 2023); industry, including sectors such as finance (Li et al. 2023), healthcare (Goyal et al. 2024) or law (Lai et al. 2024) and personal use, for example, as a personal assistant or search engine (Xiong et al. 2024; Microsoft Accessed 27/05/2025). Furthermore, the recent surge in their reasoning ability (Wei et al. 2022) and the development of cost-efficient models (Liu et al. 2024) suggest that there are still new avenues for improvement.