o5i/cxr14-bpals-trial

NIH-CXR14-BPALS是一个用于NIH ChestX-ray14数据集的标签质量审计数据集。NIH ChestX-ray14的标签是通过放射学报告自动生成的，未针对图像进行验证，因此存在大量噪声或错误（据报道准确率约为80%）。该数据集使用视觉语言模型独立重新检查每个（图像，标签）对，并返回一个信任信号：指示哪些标签得到图像良好支持，哪些是可疑、困难或可能错误标记的。它不重新标记数据，而是帮助用户识别现有标签中不可信的部分，从而将有限的专家审查资源集中在真正需要的情况下，并在训练前排除或降低不可靠标签的权重。数据集包括不同层级（试用版、标准版、高级版、企业版），提供信任分数、置信度分组、视觉一致性信号等列，并支持与原始图像数据集（BahaaEldin0/NIH-Chest-Xray-14）的集成使用。NIH-CXR14-BPALS is a label-quality audit dataset for NIH ChestX-ray14. The labels in NIH ChestX-ray14 are derived automatically from radiology reports and not verified against the images, resulting in a significant share of noisy or incorrect labels (~80% reported accuracy). This dataset independently re-examines each (image, label) pair using a vision-language model and returns a trust signal: indicating which labels are well-supported by the image and which are suspect, hard, or likely mislabeled. It does not relabel the data but instead identifies where existing labels cannot be trusted, allowing users to focus limited expert review on cases that actually need it and to exclude or down-weight unreliable labels before training. The dataset includes tiers (Trial, Standard, Premium, Enterprise) with columns such as trust score, confidence group, visual agreement signals, and supports integration with the original image dataset (BahaaEldin0/NIH-Chest-Xray-14).