danielrosehill/Whiteboards

Whiteboards数据集是一个小型、单一作者的白板图像语料库，旨在评估视觉语言模型（VLM）在手写OCR准确性上的表现，并研究伪文本幻觉（即VLM为模糊手写内容生成看似合理但错误的词语的失败模式）。所有图像均来自同一面墙挂白板、同一支马克笔和同一作者，使用手机拍摄。数据集包含21个样本，每个样本约为150 KB的WebP格式图像，分辨率为4096×3072。每个样本包含以下字段：file_name（图像路径）、id（样本ID）、category（内容类型）、transcription（手写内容的逐字转录）和description（白板内容的语义描述）。数据集主要用于零样本OCR准确性、少样本OCR、图像到图像保留、结构/语义任务以及手写特定探针等评估任务。数据集的局限性包括单一作者、单一白板、单一马克笔、无多色样本、自由形式的Markdown转录以及部分样本重叠。数据集采用CC-BY-4.0许可。

The Whiteboards dataset is a small, single-author whiteboard corpus designed for evaluating vision-language models (VLMs) on handwritten OCR accuracy and for studying pseudotext hallucination—the failure mode where a VLM invents plausible-but-wrong words for ambiguous handwriting. Every image is of the same wall-mounted whiteboard, same marker, and same author, photographed with a phone. The dataset contains 21 samples, each approximately 150 KB in WebP format, with a resolution of 4096×3072. Each sample includes fields such as file_name (path to the image), id (sample ID), category (content type), transcription (verbatim ground truth of the handwritten content), and description (prose description of the boards content). The dataset is intended for tasks like zero-shot OCR accuracy, few-shot OCR, image-to-image preservation, structure/semantics tasks, and handwriting-specific probes. Limitations include single author, single board, single marker, no multi-color samples, free-form markdown transcriptions, and some overlapping samples. The dataset is licensed under CC-BY-4.0.