srpone/zooclaw-fashion-eval
收藏资源简介:
ZooClaw-Fashion是一个用于时尚图文检索的评估基准,旨在严格测试跨模态检索模型在真实世界电商时尚产品上的性能。它包含零样本和领域内两种查询划分,支持对模型泛化能力的细粒度分析。数据来源于zoodata.ai提供的多品牌时尚目录,该平台是ZooClaw平台上智能体使用的数据代理栈。数据集包含2000个查询(1000个零样本查询和1000个领域内查询)和12000个产品,覆盖2086个独特品牌,支持文本到图像、图像到文本和文本到文本三种检索任务。数据按类别(如上衣、下装、包袋等)和人口统计属性(如女性、男性、儿童等)分布,图像格式包括WEBP、JPG和PNG。数据集结构遵循BEIR风格,包含独立的查询、语料库和真实标签文件,用于评估时尚检索模型的性能。
ZooClaw-Fashion is an evaluation benchmark for fashion image-text retrieval, designed to rigorously test cross-modal retrieval models on real-world e-commerce fashion products. It features both zero-shot and in-domain query splits, enabling fine-grained analysis of model generalization. Products are sourced from the multi-brand fashion catalog provided by zoodata.ai — the data-agent stack used by agents on the ZooClaw platform. The dataset includes 2,000 queries (1,000 zero-shot and 1,000 in-domain) and 12,000 products, covering 2,086 unique brands, and supports text-to-image, image-to-text, and text-to-text retrieval tasks. Data is distributed across categories (e.g., top, bottom, bag) and demographics (e.g., women, men, kids), with image formats including WEBP, JPG, and PNG. The dataset follows a BEIR-style structure with separate files for queries, corpus, and ground truth, facilitating evaluation of fashion retrieval models.




