proc1v/screen2ax-tree-silver-qwen3vl-235b
收藏资源简介:
Screen2AX-Tree Silver (Qwen3-VL-235B) 是一个银级(伪标记)可访问性树数据集,用于macOS屏幕截图。该数据集由Qwen/Qwen3-VL-235B-A22B-Instruct模型在零样本设置下生成,旨在用于小型视觉语言模型的知识蒸馏,以及研究大型前沿视觉语言模型在macOS GUI理解上的失败模式。数据集包含1127个样本,每个样本包含原始截图、人工标注的真实AX树和模型预测的AX树。数据集结构包括多个字段,如id、image、ax_tree_gt、ax_tree_pred等。数据集的使用场景包括蒸馏/伪标签SFT、错误分析和基准测试预热数据。数据集的限制包括坐标缩放、生成器质量参差不齐、截断和替代代码点等问题。
Screen2AX-Tree Silver (Qwen3-VL-235B) is a silver (pseudo-labelled) accessibility tree dataset for macOS screenshots. The dataset is generated by the Qwen/Qwen3-VL-235B-A22B-Instruct model in a zero-shot setting and is intended for knowledge distillation of smaller vision-language models and for studying the failure modes of large frontier VLMs on macOS GUI understanding. The dataset contains 1127 samples, each including the original screenshot, the human-curated ground-truth AX tree, and the AX tree predicted by the model. The dataset structure includes various fields such as id, image, ax_tree_gt, ax_tree_pred, etc. Intended uses include distillation/pseudo-label SFT, error analysis, and benchmark warm-up data. Limitations include coordinate scaling, mixed generator quality, truncation, and surrogate codepoints.




