I might be wrong, but fine-tuning a model based on a particular human’s behavioural output should do the job to some extent, right?