Mitigating LLM biases toward spurious social contexts using direct preference optimization