ZGZzz/NavGPT-Instruct
收藏Hugging Face2024-07-20 更新2024-07-22 收录
下载链接：
https://hf-mirror.com/datasets/ZGZzz/NavGPT-Instruct
下载链接
链接失效反馈官方服务：
资源简介：
NavGPT-Instruct-10k数据集是为了训练具有导航推理能力的视觉语言模型（VLM）而创建的。该数据集通过GPT-4V自动生成，考虑了在中间步骤中代理的导航情况。数据集从R2R训练集中随机选取了10k个中间步骤，使用等距柱状投影的全景图像作为GPT-4V的输入，以描述当前环境并指定下一步的方向或动作。
The NavGPT-Instruct-10k dataset is designed for training Large Vision-Language Models (VLM) with navigational reasoning ability. Using GPT-4V for automatic data generation, the dataset focuses on navigational reasoning at intermediate steps without considering historical modeling. It randomly selects 10,000 intermediate steps from the trajectory in the R2R training set, using the equirectangular projected panoramic image centered at the agents heading direction as the image input for GPT-4V. Each steps navigation reasoning trace describes the current environment and specifies the direction or action to proceed.
提供机构：
ZGZzz
