Tuumailubotti is a Finnish-speaking chatbot for workplace reflection designed in collaboration with The Finnish Broadcasting Company (YLE). The training dataset has been gathered from YLE employees and members of neurominorities from outside the company. The contributors were asked to imagine what type of conversations they could never have with a human HR reprsentative and write down these fictitious exchanges.
A technical challenge in representing minorities in generative AI systems is the lack data. In the Tuumailubotti project, we investigate whether it is possible to overcome this limitation using a technique called parameter-efficient finetuning (PEFT). Are we able to neuroqueer the underlying larger foundation model with a small dataset and make the chatbot manifest neurodiverse rhetoric?