PatronusAI/world_model_corpus

该数据集包含一组生成的轨迹，专为基于文本的世界建模任务设计，用于论文《Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL》。数据集包含来自九个不同环境的轨迹：Tau2Bench、SWE-Smith、DeepresearchQA、Openresearcher、Gorilla/BFCLv4、Webshop、Toolathlon、Pandora和Coderforge。训练集包含239,403个唯一的接地状态-动作轨迹，验证集有19,454个轨迹，测试集（域内、分层）有20,847个轨迹。轨迹涵盖软件工程、研究自动化、客户服务和通用工具使用环境。每个轨迹被分解为五个接地组件：初始环境状态、任务上下文、工具模式、领域规则和引导指令，将世界建模任务形式化为可引导的转移动态问题。为了促进多样性和广泛的行为覆盖，轨迹使用十二个前沿模型系列生成，包括Qwen-3.5（4B、9B、397B-A17B-FP8）、GPT-5.4（迷你版和完整版）、GPT-OSS-120B、Claude-4.6-Sonnet、Claude-4.5-Haiku、Gemini-3-Pro、MiniMax-M2.1、GLM-5和DeepSeek-V3.2。The dataset comprises 239,403 unique grounded state–action trajectories in the training split, 19,454 validation trajectories, and 20,847 held-out, stratified in-domain test trajectories. Trajectories span software engineering, research automation, customer service, and general-purpose tool use environments. Each trajectory is decomposed into five grounded components: initial environment state, task context, tool schemas, domain rules, and steering directives, formalizing the world modeling task as a steerable transition-dynamics problem. To encourage diversity and broad behavioral coverage, trajectories were generated using twelve frontier model families including Qwen-3.5 (4B, 9B, 397B-A17B-FP8), GPT-5.4 (mini and full), GPT-OSS-120B, Claude-4.6-Sonnet, Claude-4.5-Haiku, Gemini-3-Pro, MiniMax-M2.1, GLM-5, and DeepSeek-V3.2.