Dataset for BIM-Context-Aware Evidence-Traceable Safety Review Using Accident Narratives

Description

This dataset provides de-identified structured accident-narrative data and supporting prototype materials for BIM-context-aware, evidence-traceable safety review. The dataset was derived from publicly available OSHA Severe Injury Report narratives and was processed to convert unstructured accident descriptions into structured safety fields that can support BIM-based safety reasoning, evidence retrieval, and safety advisory generation. The shared package includes: (1) a de-identified master structured accident dataset; (2) a de-identified researcher-reviewed validation subset; (3) a BIM object risk profile; (4) an internal validation summary report; (5) prototype examples showing BIM context extraction, augmented evidence packaging, and generated safety advisory output; and (6) cleaned workflow notebooks for label validation, risk-profile generation, and context-aware evidence generation. The core structured fields include BIM-related object classes, construction activity labels, work-phase information, hazard triggers, injury outcomes, OSHA reference identifiers, and de-identified accident narratives. Direct identifying fields, including employer names, address fields, geographic coordinates, inspection identifiers, and internal personal paths, were removed before publication. The example BIM context files use a generic sample structural model, masked element identifiers, and masked X/Y coordinates. The materials are intended to support transparent reuse of the structured dataset, validation subset, and prototype workflow. The dataset can be used for safety knowledge structuring, BIM-linked risk review, accident evidence retrieval, safety decision support research, and reproducibility checks for the accompanying study. The generated safety advisory example is provided only as a prototype output and should not be interpreted as a certified safety plan or regulatory compliance document.

Steps to reproduce

1. Download the complete dataset package and preserve the folder structure described in 00_README.md. 2. Open 00_Data_Dictionary.xlsx to review the file inventory, column definitions, de-identification log, taxonomy descriptions, and usage notes. 3. Use 01_Dataset/Final_Master_Ensemble_v18_Success_Deidentified.csv as the main structured accident dataset. This file contains de-identified accident narratives and structured safety labels, including BIM object, work phase, action, hazard trigger, and injury-related fields. 4. Use 01_Dataset/Gemini_Final_Validation_400_Deidentified.csv as the researcher-reviewed validation subset. This file can be used to inspect label agreement and compare generated labels with reviewed reference labels. 5. Use 01_Dataset/OSC_BIM_Risk_Profile.csv to reproduce the object-level risk profile based on frequency, average severity, and risk index. 6. Review 02_Internal_Validation/OSC_Internal_Validation_Report.md for implementation-level validation results, including data readiness, retrieval sanity checks, context-weighted re-ranking checks, and evidence traceability checks. 7. Review 03_RAG_Prototype_Examples/Revit_Context_Payload_Example.json to inspect the example BIM context payload. The example contains a generic sample structural model, a masked element identifier, masked X/Y coordinates, Z-level information, and the associated safety label. 8. Review 03_RAG_Prototype_Examples/Augmented_Safety_Context_Example.json to inspect how BIM context and retrieved accident evidence were packaged for safety reasoning. 9. Review 03_RAG_Prototype_Examples/Final_OSHA_JSA_Report_Example.md as an example generated safety advisory output. This file is a prototype output and should not be used as a certified safety plan. 10. The notebooks in 04_Code are provided as cleaned workflow examples. Before execution, adjust local file paths according to the downloaded folder structure and install the required Python packages. API keys are not included in the shared files. If an external language-model API is used, configure the API key locally through a secure environment variable or local runtime setting.

Institutions

• Ewha Womans University

  Seoul, Seoul

Categories

Licence