JACKYS999/Edge-Agent-Reasoning-WebSearch-260K

Edge-Agent-Reasoning-WebSearch-260K数据集是一个大规模、合成专家工程的数据集，包含超过7亿个标记，旨在训练小型本地模型和边缘部署代理在高级问题解构和自我意识推理方面。该数据集训练模型作为预备路由器或系统2思维代理，当面对复杂、领域特定的指令时，代理会系统分解请求、识别知识差距、制定具体模糊性并构建专家级网络搜索查询。这种预备推理为后续更强大的前沿模型提供了执行最终任务所需的精确验证上下文。数据集包含260,293行，格式为Parquet，特征包括batch_index_id、role、industry、os、user_prompt和agent_reasoning。它支持推理微调、自我意识与谦逊、搜索查询生成和提示拦截等核心能力，并基于5阶段推理结构（理解请求、知识信念与不确定性、请求中的模糊性、确认前需验证内容、网络搜索查询）构建。数据集通过7维组合矩阵采样，确保高零样本多样性，覆盖200多个专业角色和多种操作系统环境，适用于分布式代理架构中的模型训练。

The Edge-Agent-Reasoning-WebSearch-260K dataset is a massive, synthetically expert-engineered corpus of over 700 million tokens, designed to train small, local models and edge-deployed agents in advanced problem deconstruction and self-aware reasoning. It trains models to act as a preparatory router or System 2 thinking agent, systematically breaking down complex, domain-specific requests, identifying knowledge gaps, formulating specific ambiguities, and constructing expert-level web search queries. This preparatory reasoning equips a secondary, more capable frontier model with the exact verified context needed to execute the final task flawlessly. The dataset contains 260,293 rows in Parquet format, with features including batch_index_id, role, industry, os, user_prompt, and agent_reasoning. It supports core capabilities such as reasoning fine-tuning, self-awareness and humility, search query generation, and prompt interception, structured around a 5-stage reasoning process (understanding the request, knowledge beliefs and uncertainties, ambiguities in the request, verification checklist, and web search queries). Sampled from a 7-dimensional combinatorial matrix, it ensures high zero-shot diversity, covering over 200 professional roles and multiple operating system environments, suitable for model training in distributed agentic architectures.