OpenMOSS-Team/VehicleWorld

--- license: apache-2.0 language: - en pretty_name: "VehicleWorld Dataset" tags: - code - simulation - autonomous-driving - large-language-models # --- Crucial for driving the dataset viewer --- dataset_info: features: - name: id dtype: string - name: execute.py dtype: string - name: inits dtype: string - name: modules dtype: string - name: querys dtype: string - name: raw dtype: string - name: worlds.json dtype: string - name: scenario_type dtype: string - name: device_type dtype: string splits: - name: train num_bytes: 155973757 num_examples: 1291 download_size: 155973757 dataset_size: 155973757 --- # 📚 Introduction **VehicleWorld** is the first comprehensive multi-device environment for intelligent vehicle interaction that accurately models the complex, interconnected systems in modern cockpits. This environment enables precise evaluation of agent behaviors by providing real-time state information during execution. This dataset is specifically designed to evaluate the capabilities of Large Language Models (LLMs) as in-car intelligent assistants in understanding and executing complex commands within realistic vehicle environments. - **GitHub Repository**: [OpenMOSS/VehicleWorld](https://github.com/OpenMOSS/VehicleWorld) - **Paper**: [VehicleWorld: A Highly Integrated Multi-Device Environment for Intelligent Vehicle Interaction (arXiv:2509.06736)](https://arxiv.org/abs/2509.06736) # 🌟 Dataset Overview Through a meticulous construction process and expert review, our final dataset contains **1,291 carefully crafted tasks**. These tasks are strategically categorized into four splits based on the primary types of devices they involve: - **Multimedia**: Tasks involving entertainment systems, audio controls, and media interfaces - **Touch Control**: Tasks requiring interaction with touchscreen displays and control panels - **Car Control**: Tasks related to vehicle operation, driving functions, and mechanical systems - **Light**: Tasks involving interior and exterior lighting systems The core strength of this dataset lies in evaluating agent performance within a highly complex environment characterized by: - 🔧 **30** comprehensive vehicle function modules - 🛠️ **250** functional APIs covering all major vehicle systems - 📊 **680** queryable state properties for complete environment awareness - ⚡ **Real-time** state feedback and execution monitoring # 🚀 Getting Started Follow these steps to integrate this dataset with the VehicleWorld project environment: ## 1. Prerequisites Clone the main VehicleWorld repository: ```bash git clone https://github.com/OpenMOSS/VehicleWorld.git cd VehicleWorld ``` ## 2. Download the Dataset Navigate to the database directory and clone the dataset: ```bash cd vehicleworld/database git clone https://huggingface.co/datasets/OpenMOSS/VehicleWorld ``` ## 3. Convert Dataset Structure Run the conversion script to organize the dataset into the required directory structure: ```bash python hf_convert.py ``` After the script completes, the task files will be properly organized in `vehicleworld/database/tasks/`, ready for use with the project's evaluation pipeline. # 📂 Dataset Structure Each sample in the dataset contains the following components: | Field Name | Description | | :------------- | :------------------------------------------------------------------------------------------------------ | | `id` | Globally unique task identifier used as the directory name | | `execute.py` | Python script for generating the corresponding `worlds.json` file | | `inits` | Python code snippet for initializing the initial environment state | | `modules` | Relevant vehicle function modules involved in the current task | | `querys` | Natural language command or question posed by the user to the in-car assistant | | `raw` | Raw text describing the task scenario, generated by the Claude-3.7-Sonnet model | | `worlds.json` | List of world state transitions during task execution | | `scenario_type`| Scenario interaction type label (`S-S`, `S-M`, `M-S`, `M-M`) derived from query/API-call structure | | `device_type` | Device category label (`Multimedia`, `Touch Control`, `Car Control`, `Light`) derived from module names | # 📚 Citation If you find VehicleWorld or this dataset useful in your research, please cite our paper: ```bibtex @misc{yang2025vehicleworldhighlyintegratedmultidevice, title={VehicleWorld: A Highly Integrated Multi-Device Environment for Intelligent Vehicle Interaction}, author={Jie Yang and Jiajun Chen and Zhangyue Yin and Shuo Chen and Yuxin Wang and Yiran Guo and Yuan Li and Yining Zheng and Xuanjing Huang and Xipeng Qiu}, year={2025}, eprint={2509.06736}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2509.06736} } ```许可证：Apache-2.0 语言： - 英语 展示名称："VehicleWorld 数据集" 标签： - 代码 - 仿真 - 自动驾驶 - 大语言模型 # --- 驱动数据集查看器的核心配置 --- dataset_info: 特征: - 名称: id 数据类型: 字符串 - 名称: execute.py 数据类型: 字符串 - 名称: inits 数据类型: 字符串 - 名称: modules 数据类型: 字符串 - 名称: querys 数据类型: 字符串 - 名称: raw 数据类型: 字符串 - 名称: worlds.json 数据类型: 字符串 - 名称: scenario_type 数据类型: 字符串 - 名称: device_type 数据类型: 字符串 划分: - 名称: train 字节数: 155973757 样本数: 1291 下载大小: 155973757 数据集大小: 155973757 --- # 📚 简介 **VehicleWorld** 是首个面向智能车辆交互的综合性多设备环境，可精准建模现代座舱内复杂且相互关联的系统。该环境可在执行过程中提供实时状态信息，从而实现对智能体行为的精准评估。本数据集专为评估大语言模型（Large Language Models，LLMs）作为车载智能助手的能力而设计，用于测试其在真实车辆环境中理解并执行复杂指令的能力。 - **GitHub 仓库**：[OpenMOSS/VehicleWorld](https://github.com/OpenMOSS/VehicleWorld) - **相关论文**：[VehicleWorld: 面向智能车辆交互的高度集成多设备环境（arXiv:2509.06736）](https://arxiv.org/abs/2509.06736) # 🌟 数据集概览 经过严谨的构建流程与专家评审，本数据集最终包含**1291个精心设计的任务**。这些任务根据其所涉及的主要设备类型，被划分为四个子集： - **多媒体类**：涉及娱乐系统、音频控制与媒体界面的任务 - **触控控制类**：需要与触控显示屏及控制面板交互的任务 - **车辆控制类**：与车辆操作、驾驶功能及机械系统相关的任务 - **灯光类**：涉及车内与车外照明系统的任务 本数据集的核心优势在于可在以下高度复杂的环境中评估智能体性能： - 🔧 **30个**完整的车辆功能模块 - 🛠️ **250个**覆盖所有主要车辆系统的功能API - 📊 **680个**可查询的状态属性，实现全环境感知 - ⚡ **实时**状态反馈与执行监控 # 🚀 快速入门 请按照以下步骤将本数据集与 VehicleWorld 项目环境集成： ## 1. 前置准备 克隆主项目仓库： bash git clone https://github.com/OpenMOSS/VehicleWorld.git cd VehicleWorld ## 2. 下载数据集 进入数据库目录并克隆本数据集： bash cd vehicleworld/database git clone https://huggingface.co/datasets/OpenMOSS/VehicleWorld ## 3. 转换数据集结构 运行转换脚本，将数据集整理为所需的目录结构： bash python hf_convert.py 脚本执行完成后，任务文件将被妥善组织至 `vehicleworld/database/tasks/` 目录下，即可用于项目的评估流程。 # 📂 数据集结构 本数据集的每个样本包含以下组件： | 字段名 | 描述 | | :------------- | :------------------------------------------------------------------------------------------------------ | | `id` | 全局唯一的任务标识符，同时作为任务所在目录的名称 | | `execute.py` | 用于生成对应 `worlds.json` 文件的 Python 脚本 | | `inits` | 用于初始化初始环境状态的 Python 代码片段 | | `modules` | 当前任务所涉及的相关车辆功能模块 | | `querys` | 用户向车载智能助手提出的自然语言指令或问题 | | `raw` | 由 Claude-3.7-Sonnet 模型生成的任务场景原始文本 | | `worlds.json` | 任务执行过程中的世界状态转换列表 | | `scenario_type`| 基于指令/API调用结构得出的场景交互类型标签（`S-S`、`S-M`、`M-S`、`M-M`） | | `device_type` | 基于模块名称得出的设备类别标签（`Multimedia`（多媒体类）、`Touch Control`（触控控制类）、`Car Control`（车辆控制类）、`Light`（灯光类）） | # 📚 引用说明 若您在研究中使用 VehicleWorld 或本数据集，请引用我们的论文： bibtex @misc{yang2025vehicleworldhighlyintegratedmultidevice, title={VehicleWorld: A Highly Integrated Multi-Device Environment for Intelligent Vehicle Interaction}, author={Jie Yang and Jiajun Chen and Zhangyue Yin and Shuo Chen and Yuxin Wang and Yiran Guo and Yuan Li and Yining Zheng and Xuanjing Huang and Xipeng Qiu}, year={2025}, eprint={2509.06736}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2509.06736} }