buildeng
收藏资源简介:
BuildEng V8 145K是一个专门为微调保守型工程助手模型而构建的土木与结构工程推理数据集。该数据集包含145,117个经过验证的聊天格式样本,核心聚焦于结构判断、检查优先推理、安全不确定性处理以及实用的建筑工程逻辑。数据集旨在训练模型进行审慎推理,而非做出最终结构批准。具体目标是教会助手模型:仔细推理、区分现场症状与根本原因、避免不安全的确定性、在需要时请求缺失信息、并在情况要求时建议合格的工程审查。数据集覆盖的工程主题广泛,包括钢筋混凝土梁、钢梁、板、柱、基础、挡土墙、砌体墙、屋顶系统、临时支撑、施工顺序、沉降、开裂、腐蚀、挠度、荷载路径变化、修复决策、矛盾的检查报告以及破坏过程。此外,还涵盖支持性建筑工程主题,如HVAC气流、管道、水力系统、电气负荷检查以及安全相关的机电管道推理。数据集训练模型遵循“观察→初步推理→缺失信息→风险检查→保守下一步行动”的行为模式。关键期望行为包括:检查优先决策、症状与原因分离、不确定性感知回答、拒绝不安全结论、矛盾处理、修复决策逻辑、荷载路径意识、土木公式推理、计算与安全限值结合、强度与使用性分离。每个样本均采用聊天消息格式,包含系统角色设定、用户问题和助手回答。该数据集主要用于土木/结构工程助手模型的监督微调,特别适用于处理不完整现场信息或潜在不安全结构条件时需要表现谨慎的模型。它已被用于BuildEng V8模型的开发,并计划用于进一步的3B和32B参数的QLoRA训练。重要安全提示:该数据集是教育和助手训练材料。基于其训练的模型不能替代持牌结构工程师、现场检查、建筑规范核查或专业工程判断。最终结构决策必须由合格专业人员根据项目具体图纸、计算、检查、标准和当地法规做出。
BuildEng V8 145K is a civil and structural engineering reasoning dataset specifically built for fine-tuning conservative engineering assistant models. It contains 145,117 validated chat-formatted samples, focusing on structural judgment, inspection-first reasoning, safety uncertainty handling, and practical construction engineering logic. The dataset aims to train models for careful reasoning rather than making final structural approvals. Specific goals include teaching assistant models to: reason carefully, distinguish between site symptoms and root causes, avoid unsafe certainties, request missing information when needed, and suggest qualified engineering review when warranted. It covers a wide range of engineering topics, including reinforced concrete beams, steel beams, slabs, columns, foundations, retaining walls, masonry walls, roof systems, temporary supports, construction sequences, settlement, cracking, corrosion, deflection, load path changes, repair decisions, conflicting inspection reports, and failure processes. Additionally, it includes supportive construction engineering topics such as HVAC airflow, plumbing, hydraulic systems, electrical load checks, and safety-related MEP reasoning. The dataset trains models to follow a behavior pattern of observation → preliminary reasoning → missing information → risk check → conservative next action. Key expected behaviors include: inspection-first decision-making, symptom-cause separation, uncertainty-aware responses, rejection of unsafe conclusions, conflict handling, repair decision logic, load path awareness, civil formula reasoning, integration of calculations with safety limits, and separation of strength from serviceability. Each sample is in chat message format with system role settings, user questions, and assistant responses. It is primarily used for supervised fine-tuning of civil/structural engineering assistant models, especially for models that need to be cautious when handling incomplete on-site information or potentially unsafe structural conditions. It has been used in the development of the BuildEng V8 model and is planned for further QLoRA training with 3B and 32B parameters. Important safety note: This dataset is educational and assistant training material. Models trained on it cannot replace licensed structural engineers, on-site inspections, building code verification, or professional engineering judgment. Final structural decisions must be made by qualified professionals based on project-specific drawings, calculations, inspections, standards, and local regulations.
数据集名称
BuildEng V8 145K
许可证
Apache-2.0
语言
英语
标签
土木工程、结构工程、建筑工程、施工、合成数据、推理
数据集规模
145,117 条经过验证的聊天格式样本
数据集聚焦领域
- 钢筋混凝土梁、钢梁、板、柱、基础、挡土墙、砌体墙、屋面系统
- 临时支撑、施工顺序、沉降、开裂、腐蚀、挠度、荷载路径变化
- 维修决策、矛盾检测报告处理、失效过程分析
- 配套建筑工程主题:暖通空调气流、风管、水暖系统、电气负载检查、与安全相关的机电管道工程推理
推理行为模式
数据集训练模型遵循以下步骤: 观察 → 初步推理 → 补充缺失信息 → 风险检查 → 保守的下一步行动
关键行为包括:
- 以检查为先的决策
- 症状与原因分离
- 不确定性感知回答
- 拒绝不安全结论
- 矛盾处理
- 维修决策逻辑
- 荷载路径感知
- 土木公式推理
- 综合计算与安全限值
- 强度与使用性分离
数据格式
每条样本使用聊天消息格式: json { "messages": [ { "role": "system", "content": "You are BuildEng, a conservative civil and structural engineering assistant." }, { "role": "user", "content": "..." }, { "role": "assistant", "content": "..." } ] }
预期用途
- 用于土木/结构工程助手的监督微调
- 特别适用于处理不完整现场信息或潜在不安全结构条件时需谨慎行为的模型
- 用于 BuildEng V8 模型开发
- 适用于进一步的 3B 和 32B QLoRA 训练
安全说明
- 数据集为教育和助手训练材料
- 基于此数据集训练的模型不能替代持证结构工程师、现场检查、建筑规范检查或专业工程判断
- 最终结构决策必须由合格专业人员根据项目具体图纸、计算、检查、标准和当地法规做出





