Liangzhi-Li/clipbench-blending

CLIPBench-Blending Code, cached features and complete per-cell records for The Blending Ratio Is Not Where the Performance Is: Diagnosing Prototype Blending for Few-Shot Adaptation of Vision–Language Models (Liangzhi Li, Bowen Wang, Yiming Qian, Thorsten Neumann, Xia Xie, and Guangshun Li; corresponding author Guangshun Li). A large family of few-shot adaptation methods for vision–language models classifies with a convex combination of the zero-shot text prototype and the mean… See the full description on the dataset page: https://huggingface.co/datasets/Liangzhi-Li/clipbench-blending.