MRiabov/IntersectionQA-15K
收藏资源简介:
IntersectionQA是一个仅包含代码的CAD空间推理基准数据集。每个示例为模型提供两个可执行的CadQuery对象构建函数和装配变换,然后要求模型推断由该代码引起的几何关系。核心问题是代码模型是否能够在心理上跟踪CAD程序的空间后果:正体积干涉、接触、接近未命中、间隙、包含和重叠量。数据集不旨在替代CAD内核。官方标签是通过CadQuery/OpenCASCADE布尔和距离查询离线生成的,然后存储在行中,以便训练和评估可以在不执行CAD代码的情况下运行。基准测试从代码中测量闭书几何基础;工具辅助的CAD代理可以单独评估作为上限设置。这是一个以基准测试为先的发布,而不是传统的以训练为主的有监督数据集。`test_near_boundary`分割特意较大,因为它是接触、接近未命中、微小重叠和反事实案例的主要诊断挑战集。IntersectionQA以两种公共大小发布:`IntersectionQA-15K`用于快速检查和烟雾实验,`IntersectionQA-90K`用于更大的基准测试/训练发布。此存储库包含来自IntersectionQA `v0.1`的`15,000`个公共任务行。
IntersectionQA is a code-only CAD spatial-reasoning benchmark. Each example gives a model two executable CadQuery object-construction functions plus assembly transforms, then asks it to infer the geometric relation induced by that code. The central question is whether a code model can mentally track the spatial consequences of CAD programs: positive-volume interference, contact, near misses, clearance, containment, and overlap magnitude. The dataset is not intended to replace a CAD kernel. Official labels are produced offline with CadQuery/OpenCASCADE Boolean and distance queries, then stored in the rows so training and evaluation can run without executing CAD code. The benchmark measures closed-book geometric grounding from code; tool-assisted CAD agents can be evaluated separately as an upper-bound setting. This is a benchmark-first release rather than a conventional train-heavy supervised dataset. The `test_near_boundary` split is intentionally large because it is the main diagnostic challenge set for touching, near-miss, tiny-overlap, and counterfactual cases. IntersectionQA is published in two public sizes: `IntersectionQA-15K` for quick inspection and smoke experiments, and `IntersectionQA-90K` for the larger benchmark/training release. This repository contains `15,000` public task rows from IntersectionQA `v0.1`.




