--- license: cc-by-4.0 task_categories: - question-answering language: - en tags: - bim - ifc - building-information-modeling - benchmark - construction size_categories: - 1K
4351
ac20
city_house_munich
dental_clinic
digital_hub
duplex
ettenheim_gis
fantasy_hotel_1
fantasy_hotel_2
fantasy_office_building_1
fantasy_office_building_2
fantasy_office_building_3
fantasy_residential_building_1
fzk_house
hitos
molio
samuel_macalister_sample_house
schependomlaan
sixty5
smiley_west
wbdg_office ## Question Taxonomy Questions are classified into 4 categories based on data retrieval complexity and computational requirements, adapted from the classification framework by [Solihin et al. (2015)](https://doi.org/10.1016/j.autcon.2015.03.004) for automated BIM compliance checking. | Category | Name | Count | Description | |----------|------|-------|-------------| | 1 | Direct Information Retrieval | 152 | Answerable through direct access to explicitly stored BIM element properties, without requiring computation. | | 2 | Computational Aggregation | 569 | Require simple, unambiguous mathematical operations (counting, summation, averaging, filtering) on directly accessible data. | | 3 | Geometric/Spatial Computation | 112 | Require complex computations involving geometric properties, spatial relationships, or derived measurements not directly stored. | | 4 | Incomplete Information | 194 | The required information is not fully available in the BIM model; answering requires assumptions or acknowledging limitations. | For details on how the taxonomy is used in evaluation, see: > Hellin, S., Nousias, S., & Borrmann, A. (2026). *Evaluating AI Systems for BIM Information Retrieval: A Multi-Dimensional Framework Beyond Binary Accuracy.* GNI 2026. ## Getting Started ### Prerequisites - Python 3.8+ - [uv](https://github.com/astral-sh/uv) (recommended for virtual environment management) ### Quick Start ```bash git clone https://github.com/sylvainHellin/ifc-bench.git cd ifc-bench # Install dependencies (using uv) uv pip install pandas ifcopenshell ``` ### Using the Dataset ```python import pandas as pd # Load V2 dataset df = pd.read_csv('questions/ifc-bench-v2.csv') print(f"Total QA pairs: {len(df)}") print(f"Projects: {df['project'].nunique()}") print(f"Category distribution:\n{df['category'].value_counts().sort_index()}") # Filter by category spatial_questions = df[df['category'] == 3] print(f"\nGeometric/Spatial questions: {len(spatial_questions)}") # Filter by project duplex_questions = df[df['project'] == 'duplex'] print(f"Duplex questions: {len(duplex_questions)}") ``` ### Dataset Columns | Column | Description | Example | |--------|-------------|---------| | `question` | Natural language question about a BIM model | "What is the total gross floor area of the building?" | | `ground_truth` | Reference answer | "The total gross floor area is 354.67 sqm" | | `ifc_model` | Model filename (without `.ifc` extension) | "arc" | | `project` | Project name (matches directory under `projects/`) | "duplex" | | `category` | Question complexity category (1-4) | 2 | ## Dataset Integrity Verify dataset integrity using SHA-256 checksums: ```bash # V2 dataset shasum -a 256 questions/ifc-bench-v2.csv # Expected: 8f08f5d04834a79310eb7de81f2d6812e74d53a01363affdb815bf86dfc4dbf4 # V1 dataset (preserved for backwards compatibility) shasum -a 256 questions/ifc-bench-v1.csv # Expected: f67a48770d74b6e0ff0868c923c3e1d976110350b2c439564d7ceccc16a46f35 ``` ## Contributing We welcome contributions: - New IFC models (with permissive licensing) - Additional QA pairs for existing models - Documentation improvements - Error corrections in existing answers Please see our [Contribution Guidelines](docs/CONTRIBUTING.md) for details. ## License - **Dataset** (question-answer pairs, documentation, model cards, and all original content): Licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). - **IFC models**: Each model retains its original license as specified in its project's `license.txt`. The majority of models are licensed under CC BY 4.0 or MIT. However, the following models are licensed under **GNU GPLv3** and are subject to its copyleft terms: - `4351`, `ettenheim_gis`, `hitos`, `samuel_macalister_sample_house` (source: [BIMserver TestFiles](https://github.com/opensourceBIM/TestFiles)) Users who redistribute or create derivative works that include these GPLv3 model files must comply with the GPLv3 license terms for those files. See each project's `license.txt` for details. ## Citation If using IFC-Bench in research, please cite the paper that introduces the v2 benchmark (currently a preprint, submitted to *Automation in Construction*; this entry will be updated once the paper is accepted): ```bibtex @misc{hellinBIMInformationExtraction2026, title = {{{BIM Information Extraction Through LLM-based Adaptive Exploration}}}, author = {Hellin, Sylvain and Jang, Suhyung and Fuchs, Stefan and Nousias, Stavros and Borrmann, Andr{\'e}}, year = 2026, eprint = {2605.01698}, archivePrefix = {arXiv}, primaryClass = {cs}, doi = {10.48550/arXiv.2605.01698}, url = {https://arxiv.org/abs/2605.01698}, note = {Preprint, submitted to Automation in Construction} } ``` The v1 benchmark was introduced in: ```bibtex @inproceedings{hellinNaturalLanguageInformation2025, title = {Natural {{Language Information Retrieval}} from {{BIM Models}}: {{An LLM-Based Agentic Workflow Approach}}}, booktitle = {{{EC3}}}, author = {Hellin, Sylvain and Nousias, Stavros and Borrmann, Andr{\'e}}, year = 2025, doi = {http://www.doi.org/10.35490/EC3.2025.265}, abstract = {While Building Information Models (BIM) effectively store building-related information, accessing it requires specialized software and expertise. Natural Language (NL) interfaces for BIM data retrieval can mitigate this challenge, but existing approaches are limited by rigid ontological frameworks or extensive pre-processing requirements. We present a Large Language Model-based agentic workflow that processes NL queries and automatically interacts with IFC-encoded BIM models without ontological or pre-processing constraints. In tests across architectural, structural, and MEP domains, our approach achieves 80\% overall accuracy. We provide open access to IFC-Bench-v1, our evaluation dataset containing various queries, answers, and reference BIM models.}, copyright = {CC0 1.0 Universal Public Domain Dedication}, langid = {english} } ``` ## Acknowledgments - [buildingSMART International](https://www.buildingsmart.org/) for providing sample files - [RWTH Aachen University E3D Institute](https://github.com/RWTH-E3D) for the Digital Hub models - [KIT IAI](https://www.ifcwiki.org/) for IFC example models - Tamira Wrabel, Sebastian Esser and Zijian Wang for preparing and curating the BIM Fundamentals SS2025 student models - TUM BIM Fundamentals SS2025 students for fantasy building models - [Selahattin Doelger](https://github.com/buildingsmart-community/Community-Sample-Test-Files) for the Sixty5 models - The openBIM community for quality assurance - Wawan Solihin (Singapore) for donating the West Riverside Hospital multi-discipline IFC models - Professor Robert Amor (University of Auckland) for hosting the models on the OpenIFC Model Repository ## Fixed evaluation split (Hellin et al. 2026) The main dataset is split-agnostic: `questions/ifc-bench-v2.csv` carries a stable `id` column (1..1027) as identity only, and defines no train/test partition. For like-for-like comparison with the results reported in Hellin et al. 2026, a fixed evaluation split is published separately at `questions/eval-split-hellin2026.csv` (columns `id,split`). It pins the exact 514 held-out test questions (by `id`) used in that paper, with the remaining 513 questions as the train set. The split is seed-free and reproducible: join it to the dataset on `id`. - Test: 514 questions. Per category (1/2/3/4): 72 / 289 / 50 / 103. - Train: 513 questions. This artifact is provided purely for reproducibility; the dataset itself remains split-agnostic, so other users are free to define their own splits over the `id` column. ### Deduplication note (v2 update) v2 now removes a duplicated question. One question ("Which door types have fire rating information?", category 2, project 4351) appeared twice, once in the train side and once in the test side of the fixed split. To avoid a train/test leak, the train-side copy (former `id` 21) was removed and the test-side copy (`id` 22) was retained. `id` values are stable identifiers, not row positions, so surviving rows keep their ids and `id` 21 is now an intentional gap. The frozen 514-question evaluation split (`questions/eval-split-hellin2026.csv`) is unchanged: the same 514 test ids (including `id` 22) are pinned, and only the leaked train row was dropped, leaving 512 train questions. The exact pre-dedup snapshot is preserved at git tag `v2.0-frozen-split`. --- **Maintainer**: Sylvain Hellin | **Contact**: [sylvain.hellin@tum.de](mailto:sylvain.hellin@tum.de) | **Issues**: [GitHub Issues](https://github.com/sylvainHellin/ifc-bench/issues)