clipbench-blending
收藏资源简介:
CLIPBench-Blending 是一个用于小样本视觉-语言模型原型混合诊断的基准数据集,旨在支持论文《The Blending Ratio Is Not Where the Performance Is》中的研究。该数据集提供了缓存的特征和完整的逐单元格实验记录,使得论文中的所有数字、表格和图表可以在几分钟内仅用 CPU 即可重现,无需重新解码图像或使用 GPU。数据内容包括:从 10 个数据集(FGVC-Aircraft、Caltech101、Stanford Cars、DTD、EuroSAT、Flowers102、Food-101、ImageNet、Oxford-IIIT Pets、SUN397)和 5 个骨干网络(OpenAI CLIP RN50、ViT-B/32、ViT-B/16、ViT-B/14、SigLIP ViT-B/16)中提取的图像特征(训练/测试分割)、文本原型(四种提示层级:照片/数据集特定/集成/CuPL)以及标签数组;支持 CoOp 协议的子集特征;部分镜像的支持集增强视图(20 个随机裁剪和翻转);以及 4800 个实验单元格×15 种方法的完整结果,包括每个单元格的拟合比率、预言值、逐类诊断和基线内部超参数。数据规模约为 5.2 GB(主要特征)、1.3 GB(CoOp 特征)、0.3 GB(结果),以及代码和清单。数据以 .npy 和 .npz 格式存储,特征已进行 ℓ2 归一化。该数据集适用于图像分类、零样本图像分类、小样本学习、视觉-语言模型基准测试和可重复性研究。
CLIPBench-Blending is a benchmark dataset for diagnosing prototype blending in few-shot vision-language models, designed to support the research in the paper "The Blending Ratio Is Not Where the Performance Is". The dataset provides cached features and complete per-cell experimental records, enabling all figures, tables, and charts in the paper to be reproduced within minutes using only CPU, without re-decoding images or using GPU. Data contents include: image features (train/test splits) extracted from 10 datasets (FGVC-Aircraft, Caltech101, Stanford Cars, DTD, EuroSAT, Flowers102, Food-101, ImageNet, Oxford-IIIT Pets, SUN397) and 5 backbone networks (OpenAI CLIP RN50, ViT-B/32, ViT-B/16, ViT-B/14, SigLIP ViT-B/16), text prototypes (four prompt levels: photo/dataset-specific/ensemble/CuPL), and label arrays; subset features supporting the CoOp protocol; partially mirrored support set augmentation views (20 random crops and flips); and complete results for 4800 experimental cells × 15 methods, including fitting ratio, oracle value, per-class diagnosis, and baseline internal hyperparameters for each cell. Data size is approximately 5.2 GB (main features), 1.3 GB (CoOp features), 0.3 GB (results), plus code and manifests. Data is stored in .npy and .npz formats, with features ℓ2 normalized. The dataset is suitable for image classification, zero-shot image classification, few-shot learning, vision-language model benchmarking, and reproducibility research.




