ONESTRUCTION Releases 'Ishigaki-IDS', a BIM Information Delivery Specification (IDS) Generation Foundation Model, and 'IDS-Bench' Evaluation Benchmark Developed via GENIAC

ONESTRUCTION Inc. (Headquarters: Tottori City, Tottori Prefecture; CEO: Daiho Nishioka), a company developing solutions for the construction sector centered on openBIM with the mission of 'Bridging Construction and Technology,' has released the 'Ishigaki-IDS' AI model and the 'IDS-Bench' evaluation benchmark. These are the results of research and development conducted as part of 'GENIAC' (Generative AI Accelerator Challenge), a project promoted by the Ministry of Economy, Trade and Industry (METI) and NEDO (New Energy and Industrial Technology Development Organization) to strengthen domestic generative AI development capabilities.

About Ishigaki-IDS and IDS-Bench

ONESTRUCTION is working on the development of 'Ishigaki,' a series of general-purpose generative AI foundation models specialized for the construction domain. As a project selected for GENIAC, we focused on the 'Research and Development of a BIM Information Requirement Generation Foundation Model for openBIM,' and are releasing the following three outcomes.

'Ishigaki-IDS': A Foundation Model for Generating Construction BIM Information Delivery Specifications (IDS)

Ishigaki-IDS is a foundation model for construction BIM, fine-tuned for IDS generation tasks based on existing open models selected for their language performance and safety evaluations. In this development, we built three model sizes: Ishigaki-IDS-8B, Ishigaki-IDS-14B, and Ishigaki-IDS-32B. Of these, we are releasing Ishigaki-IDS-8B.

Ishigaki-IDS was developed by fine-tuning Qwen3 (8B/14B/32B) specifically for IDS generation tasks. The base model, Qwen3, was selected by referencing the Nejumi Leaderboard, chosen for its proven general language performance and safety, the availability of multiple model sizes, and its widespread domestic and international adoption, which facilitates information gathering and verification. The usage model is designed so that users can download the model and complete both fine-tuning and inference within their own managed environments. Development was conducted with the assurance that training and input data are not transmitted externally.