Palliative Care Adversarial Dataset (PCAD)

This dataset is shared as part of the study titled: Large language models perpetuate bias in palliative care: development and analysis of the Palliative Care Adversarial Dataset (PCAD). Read on arXiv: https://arxiv.org/abs/2502.08073<b>Note on the dataset</b>The PCAD datasets were developed through an iterative process involving two palliative care physicians (NA and SB), a social scientist with expertise in inequity in palliative care (JD), and a clinician specialising in LLM research (FA). The following examples were used to represent axes of identity: ethnicity (White [reference] vs. Black, Asian, or Hispanic [marginalised]); age (30 years old [reference] vs. &gt;75 years old [marginalised]); and diagnosis (cancer [reference] vs. dementia, liver failure, heart failure, or chronic kidney disease [marginalised]). The complete datasets are provided as supplementary material.The PCAD-Direct dataset included 100 adversarial questions, designed to be concise, direct, and intentionally provocative, often beginning with biased premises reflecting an axis of identity. The goal was to test whether LLMs would challenge these biases. Each of the four care dimensions had 25 questions, of which 15 addressed a single axis of identity and 10 incorporated multiple axes to assess intersectionality (seven with two axes and three with three axes).The PCAD-Counterfactual dataset comprised 84 pairs of clinical scenarios, each paired with a single question based on the concept of counterfactual fairness, which assesses consistency of outcomes when protected attributes (e.g. age, ethnicity) are altered. Each pair included a reference case and a counterfactual case differing only by an axis of identity (age, ethnicity, or diagnosis). The dataset covered four care dimensions with 21 scenarios each: 15 focused on a single axis, and six addressed intersectionality (three with two axes and three with three axes).