DatologyAI/DatBench-Full
收藏资源简介:
DatBench是一个专为视觉-语言模型(VLMs)设计的评估套件,旨在提供高效、高信号的评估集合。它通过四个阶段的筛选流程构建,包括多选题到生成式问题的转换、盲解过滤、使用VLM作为裁判的质量过滤和区分性子集选择。数据集包含9个核心能力,如OCR、grounding、charts、math、spatial reasoning等,并提供了高效子集和完整套件两个变体,适用于训练循环、消融实验、快速迭代和模型比较等场景。
DatBench is a curated evaluation suite for vision–language models (VLMs) designed to be faithful, discriminative, and efficient. It is constructed via a four-stage curation pipeline applied to 33 widely used VLM benchmarks spanning 9 core capabilities (e.g., OCR, grounding, charts, math, spatial reasoning). The dataset is released in two complementary variants: a high-efficiency, high-signal evaluation set (DatBench) and a cleaned full suite (DatBench-Full), suitable for training loops, ablations, rapid iteration, and model comparison.




