3D-GRAND

该数据集包含40,087个家庭场景，与之配对的是620万密集标注的场景语言指令。这一大规模数据集能够提升3D大型语言模型的接地能力，并减少其产生的幻觉现象，同时支持从模拟到真实环境的转移。该数据集的规模达到了40,087个场景和620万条指令，其任务旨在为3D大型语言模型进行指令调优。

This dataset contains 40,087 home scenes paired with 6.2 million densely annotated scene language instructions. This large-scale dataset can enhance the grounding capability of 3D large language models, reduce their hallucination phenomena, and support sim-to-real environment transfer. With a scale of 40,087 scenes and 6.2 million instructions, this dataset is specifically designed for instruction tuning of 3D large language models.