askoepke/wit_1m_recaptioned

该数据集是从Wikipedia-based Image Text (WIT)中提取的图像-文本数据集，包含原始标题和Gemini生成的标题。数据集分为两个配置：wit_1024包含1,024个用于对齐评估的查询样本，wit_1m包含1,000,000个从WIT中抽取的样本，去除了重复的感知哈希和标题文本，并排除了1,024个查询样本。数据集列包括图像（JPEG格式）、原始标题（来自WIT）、源图像URL（Wikimedia Commons）和Gemini生成的约500字描述。所有图像均使用`gemini-3-flash-preview`模型生成标题，提示语要求生成详细的、仅基于视觉事实的描述。数据来源为Wikipedia-based Image Text (WIT)数据集（Srinivasan et al., 2021）。

Image–text dataset derived from Wikipedia-based Image Text (WIT) with original and Gemini-generated captions. The dataset includes two configs: wit_1024, a fixed set of 1,024 query samples used for alignment evaluation, and wit_1m, a gallery of 1,000,000 samples drawn from WIT, deduplicated by perceptual hash and caption text, with the 1,024 query samples excluded. Columns include image (JPEG), original_caption (from WIT), url (Wikimedia Commons), and gemini_caption (~500-word description). All images were recaptioned using `gemini-3-flash-preview` with a prompt to generate detailed, factual descriptions. Source: Wikipedia-based Image Text (WIT) dataset (Srinivasan et al., 2021).