Figureure 8 · summarizes the data usage across training and evaluationFigure 8 summarizes the data usage across training and evaluation. The main training set consists of 400K ABC models. For in-domain retrieval testing, we use a held-out ABC split of 91K CAD models. For zero-shot retrieval transfer, we evaluate on two unseen datasets: Automate with 65K models and CADParser with 40K models. This setup clearly separates indomain retrieval from zero-shot transfer evaluation, allowing us to test both memorization-free retrieval within the same CAD source and generalization to different CAD repositories. Figure 8: Overview of training, in-domain retrieval, and zero-shot retrieval splits used in our experiments.这张图概括 BRepCLIP 的整体方法流程。阅读时先看模块之间传递的训练信号，再看作者如何把目标拆成可优化的子问题。Figureure 13 · : Additional qualitative results for zero-shot classification on FabWaFigure 13: Additional qualitative results for zero-shot classification on FabWave. BRepCLIP produces more semantically accurate class predictions for engineering CAD models than point-based and multimodal baselines.这张图/表用于判断 BRepCLIP 的实验收益来自哪里。重点看替换、消融或跨模型设置下趋势是否一致，而不是只看单个最高分。
