ONESTRUCTION Releases "Ishigaki-IDS," a Foundation Model for Generating BIM Information Requirements (IDS), and "IDS-Bench," an IDS Generation Evaluation Benchmark, Developed Under GENIAC

With the mission "To be a bridge between construction and technology," ONESTRUCTION Inc. (Headquarters: Tottori City, Tottori Prefecture; CEO: Oho Nishioka), which develops solutions in the construction field centered on openBIM, has released the AI model "Ishigaki-IDS" and the evaluation benchmark "IDS-Bench," which are the results of research and development conducted as part of the "GENIAC (Generative AI Accelerator Challenge)" project promoted by the Ministry of Economy, Trade and Industry and NEDO (New Energy and Industrial Technology Development Organization) to strengthen domestic generative AI development capabilities.

About Ishigaki-IDS and IDS-Bench

ONESTRUCTION is developing the "Ishigaki" series of general-purpose generative AI foundation models specialized for the construction domain. This time, as a GENIAC-selected project, we worked on the "Research and Development of a Foundation Model for Generating BIM Information Requirements in openBIM." As a result, we are releasing the following three items.

"Ishigaki-IDS," a Foundation Model for Generating Construction BIM Information Requirements (IDS)

Ishigaki-IDS is a foundation model for construction BIM that has been additionally trained for IDS generation tasks, based on an existing open model selected after language performance and safety evaluations. In this development, we built three types of models: Ishigaki-IDS-8B, Ishigaki-IDS-14B, and Ishigaki-IDS-32B. Of these, we are releasing Ishigaki-IDS-8B.

Ishigaki-IDS was developed by performing additional training on Qwen3's 8B/14B/32B models to specialize them for the IDS generation task. The base model was selected with reference to the Nejumi Leaderboard 4. Qwen3 was chosen because it ensures a certain level of general-purpose language performance and safety, is available in multiple model sizes, and is a widely used base model both in Japan and internationally, making information gathering and verification easier. For our use case, we have adopted a configuration where the model is downloaded, and all additional training and inference are completed within a managed environment. Therefore, we have proceeded with development after confirming that no training or input data is sent externally.

"IDS-Bench," a Benchmark for Evaluating IDS Generation

IDS-Bench is a proprietary benchmark built in collaboration with experts in the construction and BIM fields to measure the generation accuracy of CSV-to-IDS. It is designed to evaluate IDS generation performance from multiple perspectives, ensuring coverage in terms of IFC/IDS version, language, and construction category.

Simultaneous Release of a Tech Blog Summarizing the Development Process

We have published details about the knowledge and technical background gained from this development on our tech blog.

※ openBIM is an approach for the collaborative design, construction, and operation in construction, combining open standards and workflows.

Details: https://note.com/onestruction/n/n9ffa7dda5d01

※ IDS (Information Delivery Specification) is a requirements document that defines "what information should be included" in BIM data. It is an approach to building construction similar to test-driven development.

Details: https://note.com/onestruction/n/n0918976da166

About the IDS Generation Performance of Ishigaki-IDS

The results of evaluating the newly developed Ishigaki-IDS with IDS-Bench confirmed its high performance in the IDS generation task. We believe the score difference compared to general-purpose models is due to the fact that IDS is a highly specialized standard in the construction BIM domain and is also a relatively new standard published in 2024, making it difficult for general-purpose pre-training alone to handle it sufficiently. Ishigaki-IDS has achieved improved IDS generation performance through additional training that takes into account such specialized knowledge and standard structures. Please refer to our tech blog for evaluation details.

Future Outlook

We plan to continue publishing the knowledge gained from this initiative in papers and tech blogs. ONESTRUCTION will continue to promote construction DX by leveraging technology, including AI.

Hiring Information

ONESTRUCTION is looking for individuals interested in construction AI and construction foundation model development. Please see the job posting for details.