openworlds
OpenWorldSandbox
OpenWorldSandbox Dataset
场景与任务数据(JSON v0.1 规范)——OpenWorldSandbox 具身场景沙盒的官方数据集。
代码仓库:https://github.com/ZetingLiu/OpenWorldSandbox(仓库只含代码与规格文档,数据托管于此)
沙盒说明:具身场景沙盒,JSON 场景/任务 → compile 为 SQLite 世界快照,17 个语义动作,部分可观测,Goal DSL 判卷。
内容
路径
说明
scenarios/
场景包:home_01(家庭,6 区 29 实体)、market_01(超市,6 区 26 实体)
tasks/
11 个任务:home 6 个(含 2 个探索任务变体对)、market 5 个(含 2 个探索任务变体对)
scenarios/README.md
场景 JSON 规格 v0.1(冻结)
tasks/README.md
任务 JSON 规格 v0.1(冻结)… See the full description on the dataset page: https://huggingface.co/datasets/zetliu2001/OpenWorldSandbox.openworlds-ad-trajectories
OpenWorlds AD Penetration Testing Trajectories
Synthetic training trajectories for fine-tuning AI agents on Active Directory penetration testing.
Dataset Description
Each trajectory is a complete attack path from initial access to Domain Admin, formatted as multi-turn chat conversations with:
System prompt with target domain context
reasoning traces (expert thought process)
Tool calls (nmap, ldapsearch, GetUserSPNs, hashcat, secretsdump, etc.)
Tool outputs (realistic… See the full description on the dataset page: https://huggingface.co/datasets/omkar6699/openworlds-ad-trajectories.
