UHR-CoZ #134
Description
📌 Overview
Total Samples: 25,467 samples
Dataset Size: 12.1GB
Source (): HighRS-VQA
📌 Data Pipeline
Constructed via a two-stage automated pipeline: In the annotation generation phase, a GLM-4.5V-powered agent produces multi-round zooming trajectories; in the quality control phase, answer cleansing (semantic consistency scoring) and trajectory cleansing (removal of invalid turns) are performed.