STEM is a class where students at Mass Academy learn research and engineering skills by working on an independent research project and an assistive technology project. This page focuses on the independent research project (for the assistive technology project, see STEM II). During the independent research project, students learn to document their project, use data analysis tools, and use laboratory materials. Students write a grant proposal and a STEM thesis for this project, which is later submitted to the MSEF/ISEF competition.
My project focuses on fighting the various societal biases that large language models (LLMs) exhibit. Previous work has demonstrated that LLMs show various biases based on race, gender, ethnicity, religion, and more. This work uses a promising new interpretability-based technique called Contrastive Activation Addition (CAA) to change an LLM's behavior and reduce it's measured scores in a bias benchmark. For this project, I use the open-source model Llama 2-Chat.