CaReBench data, CaRe models and all the contrastively trained MLLMs (including InternVL2, MiniCPM-V 2.6, LLaVA NeXT Video, Qwen2-VL and Tariser).