C-Fashion; MIT-States*
收藏资源简介:
C-Fashion是一个专为时尚领域组合推理构建的新基准数据集,填补了该领域在组合零样本学习任务中的空白。MIT-States*是对原始MIT-States数据集的清理和优化版本,解决了约70%标签错误的问题,提供了更可靠的评估基准。这两个数据集旨在支持组合零样本学习的研究,特别是在处理标签空间分布变化和长尾分布场景下的模型性能评估。
C-Fashion is a novel benchmark dataset constructed specifically for compositional reasoning in the fashion domain, filling the gap in compositional zero-shot learning tasks within this field. MIT-States* is a cleaned and optimized version of the original MIT-States dataset, which resolves approximately 70% of label errors and provides a more reliable evaluation benchmark. These two datasets are designed to support research on compositional zero-shot learning, particularly for evaluating model performance in scenarios involving label space distribution shifts and long-tailed distributions.
WARM-CAT 数据集概述
数据集简介
WARM-CAT 是一个面向组合零样本学习(Compositional Zero-Shot Learning)的研究项目,其核心是提出了一种名为“Warm-Started Test-Time Comprehensive Knowledge Accumulation”的方法。该项目发布了两个配套数据集。
发布数据集
项目包含以下两个公开发布的数据集:
-
C-Fashion 数据集
- 访问地址:https://huggingface.co/datasets/BJTUYXD/C-Fashion
-
MIT-States 数据集*
- 访问地址:https://huggingface.co/datasets/BJTUYXD/MIT-States_star
项目状态
该项目目前正在组织相关代码。



