BIPOLAR: Polarization-based granular framework for LLM bias evaluation