Send message to Author

Paper short abstract

This contribution explores the realities and imaginaries emerging within a participatory effort to construct a gender-oriented fine-tuning dataset for Large Language Models.

Paper long abstract

In the field of Natural Language Processing (NLP), fine-tuning is the process of adapting an LLM to a desired behaviour, tone or task. As such, fine-tuning constitutes one of the primary pathways for LLM alignment, defined as steering the behaviour of a model towards certain human values or preference. In this context, the field of STS invites the question: whose values and preferences get routinely represented in LLM alignment, and whose get discarded? This contribution explores the construction of a Co-Designed Gender Instruction Tuning Dataset (CoDIGIT), reflecting the preferences and situated knowledges of a group of 84 participants. Through a guided walk-through of the dataset and selected fine-tuning procedures, this contribution explore the social realities emerging from the encounter between participants' positionalities and Meta's LLaMA 3.1 8B model. Following the 'instruction tuning' paradigm, the fine-tuning dataset consists of 105 prompt-response pairs written by participants. Participants' responses are short stories that capture their imaginaries of what an 'aligned' LLM should output in the context of gender, but also offer insight into what participants imagine as feasible LLM outputs. This rich encounter between humans and machines reveals the kinds of identities that participants perceive as plausible or possible while interacting with LLMs, offering a glimpse into participatory conceptualisations of technological 'repair' and into the realities that materialise and emerge through participatory social research in Artificial Intelligence (AI) and NLP.