RESOURCE2SKILL
收藏资源简介:
RESOURCE2SKILL数据集是由微软官方发布的,旨在支持RESOURCE2SKILL系统的数据集。该系统将人类创建的多模态资源(如教程、文档等)提炼为可重用、可执行的软件代理技能。数据集包含结构化技能条目和可执行资产,覆盖Web、PowerPoint、Excel、Blender和REAPER风格的音频工作流。具体内容包括:skills_wiki/目录提供用于发现和检查的结构化技能条目;skills_library/目录包含可执行资产、辅助模块和领域资源;skills-lock.json文件锁定发布的技能集;CITATION.cff文件提供引用元数据。技能条目在可用时保留来源和出处元数据,确保可追溯性。数据集规模中等,样本数量在1,000到10,000之间,适用于构建软件代理的技能库、多模态资源处理、自动化工作流生成等任务。数据集采用MIT许可证,语言为英语,支持代理、技能库、多模态等应用场景。
The RESOURCE2SKILL dataset, officially released by Microsoft, is designed to support the RESOURCE2SKILL system. This system distills human-created multimodal resources (such as tutorials, documents, and other similar materials) into reusable and executable software agent skills. The dataset contains structured skill entries and executable assets, covering workflows for Web, PowerPoint, Excel, Blender, and REAPER-style audio editing. Its specific contents are as follows: the skills_wiki/ directory provides structured skill entries for discovery and inspection; the skills_library/ directory houses executable assets, auxiliary modules, and domain-specific resources; the skills-lock.json file locks the published skill collection; and the CITATION.cff file provides citation metadata. Skill entries retain source and provenance metadata whenever available to ensure traceability. The dataset is of medium scale, with the number of samples ranging from 1,000 to 10,000, and is applicable to tasks such as building software agent skill libraries, multimodal resource processing, and automated workflow generation. The dataset is licensed under the MIT License, uses English as its primary language, and supports application scenarios including AI agents, skill libraries, and multimodality.
数据集概述
数据集名称:Resource2Skill: Executable Agent Skill Libraries
发布方:Microsoft(微软)
许可协议:MIT License
语言:英语(en)
数据集规模:1K < n < 10K(样本数量在1000至10000之间)
标签:agents, skill-library, multimodal, powerpoint, web, excel, blender, audio
数据集简介
Resource2Skill 是一个由微软发布的数据集,用于将人类创建的多模态资源蒸馏为软件智能体可重用的可执行技能。该数据集是 Resource2Skill 系统的官方配套数据。
- 项目主页:https://microsoft.github.io/Resource2Skill/
- 论文地址:https://arxiv.org/abs/2606.29538
- 代码仓库:https://github.com/microsoft/Resource2Skill
数据集内容结构
数据集包含以下核心组成部分:
- skills_wiki/:结构化技能条目,用于技能发现和浏览。
- skills_library/:可执行资产、辅助模块和领域资源。
- skills-lock.json:清单文件,用于锁定已发布的技能集合。
- CITATION.cff:引用元数据。
覆盖领域
该数据集涵盖了以下领域的技能和工作流:
- Web(网页相关)
- PowerPoint(演示文稿)
- Excel(电子表格)
- Blender(3D建模)
- REAPER-style audio workflows(音频处理工作流,类似REAPER软件风格)
技能条目在可用情况下保留了来源和出处元数据。
配置与数据文件
数据集包含一个默认配置:
- 配置名称:default
- 数据文件:
- 分割:excel_validation_examples
- 路径模式:
skills_library/excel/validation/*/skill.json
使用方式
下载数据集到 Resource2Skill 代码仓库根目录:
bash hf download microsoft/RESOURCE2SKILL --repo-type dataset --local-dir .
验证某个领域(例如 PowerPoint):
bash python cli.py domains python cli.py validate-domain --domain ppt
引用格式
如需引用该数据集,请使用以下 BibTeX 格式:
bibtex @misc{fan2026resource2skill, title = {{RESOURCE2SKILL}: Distilling Executable Agent Skills from Human-Created Multimodal Resources}, author = {Yijia Fan and Zonglin Di and Zimo Wen and Yifan Yang and Mingxi Cheng and Qi Dai and Bei Liu and Kai Qiu and Yue Dong and Ji Li and Chong Luo}, year = {2026}, eprint = {2606.29538}, archivePrefix = {arXiv}, primaryClass = {cs.SE}, url = {https://arxiv.org/abs/2606.29538} }




