RefRad2D is a large-scale bilingual dataset of 1.2M CT and MR image-text pairs from clinical practice. Trained on this data, RadGrounder achieves competitive results in VQA and report generation while maintaining language quality through spatial grounding supervision without performance degradation.