orpheon-structural-dataset
收藏资源简介:
ORPHEON Structural Dataset v4 是一个专注于结构工程领域的大规模文本生成数据集,旨在支持人工智能模型在建筑和工程任务中的微调与应用。数据集包含 264,494 个示例,覆盖西班牙语、英语、葡萄牙语、罗马尼亚语和法语五种语言,确保 100% 的语言一致性。数据内容围绕 17 个工程工具(如 search_web、get_normative、lookup_section、verify_bending、design_rc_section 等)的调用展开,涉及 48 个工程类别,包括结构设计、验证、规范查询和项目管理等。v4 版本引入了关键改进:新增 132 个独特的对抗性回答以增强鲁棒性;扩展工具集至 17 个,涵盖网络搜索、规范获取、代理运行、项目验证和计算等功能;木材相关数据基于真实场景,其中 57% 通过验证,15% 失败;完全消除人工类别,确保数据真实性;并将旧工具 query_bim_element 替换为 lookup_section。数据集经过严格验证:所有工具调用均针对真实代码进行 100% 验证,质量分数不低于 0.7,采用多来源交叉验证和 1000 多个独特模板以防止过拟合。数据文件包括完整的组合数据集及标准的训练、验证和测试分割(分别为 211,595、26,449 和 26,450 条记录)。该数据集适用于结构工程相关的文本生成任务,如基于自然语言的工具调用、规范遵循、设计辅助和项目分析,支持多语言环境下的 AI 模型开发与评估。
ORPHEON Structural Dataset v4 is a large-scale text generation dataset focused on the structural engineering domain, designed to support the fine-tuning and application of artificial intelligence models in construction and engineering tasks. The dataset contains 264,494 examples, covering five languages: Spanish, English, Portuguese, Romanian, and French, ensuring 100% language consistency. The content revolves around the invocation of 17 engineering tools (such as search_web, get_normative, lookup_section, verify_bending, design_rc_section, etc.), involving 48 engineering categories including structural design, verification, normative querying, and project management. Version v4 introduces key improvements: addition of 132 unique adversarial responses to enhance robustness; expansion of the toolset to 17, covering functions like web search, normative retrieval, agent execution, project verification, and calculations; wood-related data is based on real scenarios, with 57% validated and 15% failing; complete elimination of artificial categories to ensure data authenticity; and replacement of the old tool query_bim_element with lookup_section. The dataset undergoes rigorous validation: all tool invocations are 100% validated against real code, with quality scores no lower than 0.7, employing multi-source cross-validation and over 1,000 unique templates to prevent overfitting. Data files include the complete combined dataset and standard training, validation, and test splits (211,595, 26,449, and 26,450 records, respectively). This dataset is suitable for text generation tasks related to structural engineering, such as natural language-based tool invocation, normative compliance, design assistance, and project analysis, supporting AI model development and evaluation in multilingual environments.
🏗️ ORPHEON 结构数据集 v4 概述
基本信息
- 许可证: MIT
- 语言: 西班牙语、英语、葡萄牙语、罗马尼亚语
- 数据量: 约 264,494 个样本(介于 100K 至 1M 之间)
- 任务类别: 文本生成
数据集内容
- 工具数: 17 种结构工程相关工具
- 类别数: 48 个类别
- 包含工具列表:
- search_web, get_normative, run_agent, validate_project, get_project_summary, get_project_quantities, calculate_duration_days, lookup_section, verify_bending, design_rc_section, verify_timber_bending, verify_bolted_connection, get_seismic_params, verify_shear, classify_section, check_deflection, calculate_footing
版本更新亮点 (v4)
- 新增 132 个独特对抗性回答(之前为 2 个)
- 增加 7 个 MCP 工具(总数从 11 个增至 17 个)
- 100% 语言一致性(支持 ES/EN/PT/RO)
- Timber 真实模拟:57% 通过,15% 失败
- 无人工类别(从 2000 个精确类别中移除)
- 工具替换:
query_bim_element已被lookup_section替代
质量验证
- 100% 工具调用经真实代码验证
- 质量评分 ≥ 0.7
- 三重交叉验证(3 个以上来源)
- 抗过拟合:1000+ 个独特模板
数据文件划分
| 文件 | 记录数 |
|---|---|
| data/final_dataset_combined.jsonl | 264,494 |
| data/final_dataset_train.jsonl | 211,595 |
| data/final_dataset_val.jsonl | 26,449 |
| data/final_dataset_test.jsonl | 26,450 |
使用示例
python from datasets import load_dataset dataset = load_dataset("rapman/orpheon-structural-dataset", split="train")
适用领域
- 结构工程
- 建筑规范(Eurocode、AISC、BIM)
- 微调(fine-tuning)





