Fine-Tuning as Repair? Care Ethics and Situated Knowledges in LLM Alignment Cultures

We explore the current culture of alignment and evaluation in Large Language Models (LLMs) through a novel interdisciplinary fine-tuning pipeline that centers participant expertise, situated knowledges, and subjective experiences of LLM interactions. We demonstrate proof of concept through an experimental implementation focused on LLMs' gendered biases. Through fine-tuning LLaMA 3.1 8B on a participatory dataset curated with 65 participants, and evaluating the fine-tuned model for gender bias with another 19 participants in an interactive fashion, we posit that it is possible to conceptualise fine-tuning as an act of `repair' when grounded in principles of care, consent and agency.