As is usually the case, check the data! A lot of the dataset used has fairly morbid scenarios, so the model is working as expected. All the data was synthetically created with GPT4