README.md

metadata

pretty_name: Ishigaki-Bench
license: cc-by-4.0
configs:
  - config_name: IFC2X3
    default: true
    data_files: metadata/IFC2X3.json
  - config_name: IFC4
    data_files: metadata/IFC4.json
  - config_name: IFC4X3_ADD2
    data_files: metadata/IFC4X3_ADD2.json

This result was produced as part of the "GENIAC (Generative AI Accelerator Challenge) Project" which is promoted by the Ministry of Economy, Trade and Industry and the New Energy and Industrial Technology Development Organization (NEDO), with the aim of strengthening Japan’s capabilities in generative AI development.

This dataset is an evaluation dataset for the CSV-to-IDS task used to evaluate the Ishigaki-IDS model. It evaluates whether an LLM can generate appropriate IDS from CSV files used for IFC data checking.

[Coming Soon] A high-difficulty benchmark created by experts and grounded more closely in real use cases is also planned for release at a later date.

• The technical blog covering the model overview, training, and evaluation is available here
• The released model is available here

Features

• Created in collaboration with BIM/IDS experts
• Enables systematic evaluation across three axes: understanding of IDS, IFC, and Japanese/English
• Covers four construction domains: Common, Architectural, Structural, and Equipment
• Supports three versions: IFC2X3, IFC4, and IFC4X3_ADD2

Data Specification

This dataset was created by expanding 100 base items across three IFC versions: IFC2X3, IFC4, and IFC4X3_ADD2. Each version has the same structure and is designed so that the distribution of categories and conditions is as even as possible. As a result, the dataset contains 100 samples per IFC version, for a total of 300 samples overall.

{
  "id": "001",
  "entity": "IFCPROJECT",
  "category": "Common",
  "language": "EN",
  "ids_header": true,
  "ifc_content": true,
  "input_csv": "data/input_csv/IFC2X3/IFC2X3_001_IFCPROJECT.csv",
  "output_ids": "data/output_ids/IFC2X3/IFC2X3_001_IFCPROJECT.ids"
}

Description of Each Field

• id

  A unique ID for the data.

• entity

  The IFC Entity targeted by the data.

• category

  A category indicating which construction domain the content belongs to. There are four types: Common, Architectural, Structural, and Equipment.

• language

  The language used. English is represented as EN, and Japanese as JP.

• ids_header

  Indicates whether IDS vocabulary is used in the input header. If true, IDS vocabulary is used; if false, non-IDS vocabulary is used.

• ifc_content

  Indicates whether IFC vocabulary is used in the data content. If true, IFC vocabulary is used; if false, general expressions are used.

• input_csv

  The file path where the input data for the model is stored.

• output_ids

  The file path where the ground-truth IDS data is stored.

Breakdown by Version

The number of samples for each category and condition in each IFC version is as follows.

Category	ids_header	ifc_content	language	Count per version
Common	true	true	EN	4
Common	true	true	JP	4
Common	true	false	EN	3
Common	true	false	JP	3
Common	false	true	EN	3
Common	false	true	JP	3
Common	false	false	EN	3
Common	false	false	JP	3
Architectural	true	true	EN	3
Architectural	true	true	JP	3
Architectural	true	false	EN	4
Architectural	true	false	JP	3
Architectural	false	true	EN	4
Architectural	false	true	JP	3
Architectural	false	false	EN	3
Architectural	false	false	JP	3
Structural	true	true	EN	3
Structural	true	true	JP	3
Structural	true	false	EN	3
Structural	true	false	JP	3
Structural	false	true	EN	3
Structural	false	true	JP	4
Structural	false	false	EN	4
Structural	false	false	JP	3
Equipment	true	true	EN	3
Equipment	true	true	JP	3
Equipment	true	false	EN	3
Equipment	true	false	JP	3
Equipment	false	true	EN	3
Equipment	false	true	JP	2
Equipment	false	false	EN	2
Equipment	false	false	JP	3

Benchmark Score

The generated IDS is evaluated using the "IDS-Audit-Tool(ver 1.0.96)" provided by buildingSMART. Because IDS is a specialized standard in the construction and BIM domain, and also a relatively new standard released at 2024, it is an area that general foundation models have not yet been able to handle sufficiently. Even under such conditions, this model demonstrates strong performance. Please refer to the detailed evaluation results for more information.

Evaluation Method

The source code will be shared on GitHub at a later date.

Developers

• Koyo Hidaka

• Ryo Kanazawa

*In alphabetical order

License

Creative Commons Attribution 4.0 International (CC BY 4.0)

Citation

@misc{idsbench2026,
  title={ONESTRUCTION/IDS-Bench},
  author={ONESTRUCTION Inc.},
  year={2026},
  url={https://huggingface.co/ONESTRUCTION/IDS-Bench}
}