PyDebiaser
Published:
A Debiasing Library for Transformer Models
AI Safety · LLMs — Open Source
PyDebiaser was one of my undergraduate projects, developed as part of the Fatima Fellowship under the supervision of Abubakar Abid. It is a Python library for measuring and mitigating social biases in large language models.
It implements seven debiasing techniques — SentenceDebias, INLP, Self-Debias, Bias-Swapping, Character-Neutralization, Prepend-Adjective, and Top-k — and is released under the MIT license.
