data-archetype/ocr_captions

这是一个基于WebDataset风格的分桶导出数据集，源自Yesianrohn/OCR-Data，包含图像与简洁英文描述配对，用于文本到图像训练。源数据集汇总了公共OCR基准数据，包括图像、识别文本、文本区域边界框和多边形标注。此导出保留了源OCR元数据在JSON附带文件中，并将训练描述替换为专注于可见图像内容、可读文本、文本位置、颜色和字体样式的简短描述。所有OCR-Data分割（除了`UberText`）均包含在内。

This dataset is a bucketed WebDataset-style export of Yesianrohn/OCR-Data, with images paired with concise English captions for text-to-image training. The source dataset aggregates public OCR benchmarks with images, recognized text, text-region bounding boxes, and polygon annotations. This export keeps the source OCR metadata in the JSON sidecars and replaces the training captions with short descriptions focused on the visible image content, readable text, text placement, color, and lettering style. All OCR-Data splits except `UberText` are included.