radiativity/UnitSafe

UnitSafe是一个评估AI模型是否能进行维度正确计算并区分具有相同SI维度但物理性质不同的量的基准测试。它是第一个设计用于测试**量类(KOQ)区分**能力的基准——即识别扭矩≠能量、吸收剂量≠等效剂量、视在功率≠实际功率等，即使每对量具有相同的维度公式。数据集包含500个问题，分为转换问题和必须失败问题，覆盖13个科学领域，10个KOQ退化集群，62个独特的SI签名，102个独特的量类，以及4个难度等级。

UnitSafe evaluates whether AI models can perform dimensionally correct calculations and distinguish between physically different quantities that share identical SI dimensions. It is the first benchmark designed to test **kind-of-quantity (KOQ) discrimination** — the ability to recognize that torque ≠ energy, absorbed dose ≠ equivalent dose, and apparent power ≠ real power, even though each pair has the same dimensional formula. The dataset contains 500 problems, divided into conversion problems and must-fail problems, covering 13 scientific domains, 10 KOQ degeneracy clusters, 62 unique SI signatures, 102 unique quantity kinds, and 4 difficulty tiers.