MM-Hallu/PROVE
收藏Hugging Face2026-04-30 更新2026-05-03 收录
下载链接：
https://hf-mirror.com/datasets/MM-Hallu/PROVE
下载链接
链接失效反馈官方服务：
资源简介：
PROVE是一个用于评估视觉语言模型（VLM）自由形式回答中幻觉现象的基准测试工具，基于场景图表示。它包含10,606个QA对，这些QA对基于来自超详细DOCCI图像描述的结构化视觉属性元组。数据集的特征包括图像、图像URL、超详细图像描述、问题、真实答案、图像元组和QA差异元组。元组涵盖了颜色、形状、材料、空间关系、大小、纹理等方面。评估指标包括准确性和程序化验证。
Benchmark for evaluating hallucinations in VLM free-form responses using scene-graph representations. 10,606 QA pairs grounded in structured visual property tuples from hyper-detailed DOCCI image captions. Features include image, image URL, hyper-detailed image caption, question, ground truth answer, image tuples, and QA diff tuples. Tuples cover color, shape, material, spatial relations, size, texture, etc. Evaluation metrics include Accuracy and Programmatic verification.
提供机构：
MM-Hallu
