bsdibi/foundation-pit-engineering-instruction-dataset

# Foundation Pit Engineering Instruction Dataset

## Dataset Description

This dataset is constructed to support the research presented in the paper *"A Knowledge Service System for Foundation Pit Engineering Based on LoRA and RAG(https://doi.org/10.1155/adce/8953677)"*. It consists of high-quality instruction-response pairs derived from national and industry standards for foundation pit engineering, designed for fine-tuning large language models (LLMs) in the specialized domain of geotechnical engineering.

## Dataset Summary

- **Total Samples**: 8,000+ Q&A pairs
- **Data Sources**: 35 national and industry standards
- **Language**: Chinese
- **Task Type**: Instruction Fine-Tuning (Supervised Fine-Tuning)

## Data Format

Each data sample follows the standard instruction format for supervised fine-tuning (SFT). The dataset is provided in **JSONL** format (one JSON object per line).

### Example

### Fields

| Field | Type | Description |
|-------|------|-------------|
| `instruction` | string | The user's question or instruction |
| `output` | string | The expected response from the model, based on relevant standards |

> **Note**: The `system` field is not included in this version. Users are recommended to prepend an appropriate system prompt (e.g., "You are a senior foundation pit engineering expert.") during fine-tuning.

## Data Sources

The dataset is derived from the following regulatory documents (35 in total), covering the full lifecycle of foundation pit engineering:

- JGJ 120-2012: Technical Code for Support of Building Foundation Pits
- GB 50497-2019: Technical Standard for Monitoring of Building Foundation Pit Engineering
- Other national, industrial, and local standards related to design, construction, monitoring, and maintenance.

## Data Construction Process

The dataset was constructed using a hybrid strategy combining automated generation and manual revision:

1. **Text Extraction**: PDF documents were converted to plain text using PyPDF2 and Tesseract OCR.
2. **Structured Parsing**: Text was parsed and split according to "chapter-section-article-clause" hierarchy.
3. **Q&A Generation**: GPT-4 was prompted with structured templates to generate instruction-response pairs.
4. **Quality Revision**: Automated generation outputs were manually reviewed and corrected for formula errors, internal references, and table/chart citations.

## Data Statistics

| Split | Number of Samples |
|-------|-------------------|
| Training Set | ~7,500 |
| Validation Set | ~500 |
| **Total** | **8,000+** |

## Use Cases

This dataset is intended for:

- Supervised fine-tuning (SFT) of LLMs for foundation pit engineering Q&A
- Domain adaptation of general-purpose LLMs to geotechnical engineering
- Evaluation of model performance on professional, knowledge-intensive tasks

## Limitations

- The dataset is limited to Chinese regulatory documents and may not generalize to other languages or jurisdictions.
- Some answers may reference tables or figures that are not included in the text; users should refer to the original standards for complete information.
- The dataset has been manually revised but may still contain residual errors.

## License

This dataset is released under the **MIT License**. You are free to use, modify, and distribute it, provided that proper attribution is given to the original source.

## Citation

If you use this dataset in your research, please cite the following paper:

## Contact

For questions, suggestions, or issues regarding this dataset, please contact:

- Shuqin Shi: 2422191019@xmut.edu.cn
- Chenfei Wang (Corresponding Author): wcf001@outlook.com

## Acknowledgements

This work is supported by the Scientific Research Fund of the School of Civil Engineering and Architecture, Xiamen University of Technology (Grant No. 4030226004).