How ONESTRUCTION built the Ishigaki-IDS foundation model with AWS GenAIIC
Amazon Web Services Koyo Hidaka
ONESTRUCTION built an AI model for IDS in construction with AWS help. It used synthetic data and audits to make niche BIM tasks usable for non-specialists.
Based on reporting by Amazon Web Services, Koyo Hidaka — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
ONESTRUCTION and AWS Japan say they built Ishigaki-IDS, a foundation model aimed at a very specific job: handling IDS work inside BIM workflows. The project sat inside GENIAC Phase 3, with technical advisory from the AWS Generative AI Innovation Center. That detail matters because this is exactly the sort of problem general models tend to stumble over — a narrow standard, little public data, and output that has to be structurally right, not just fluent.
The company’s target was construction, where BIM is being pushed nationally in Japan but still carries a steep learning curve. IDS, or Information Delivery Specifications, adds another layer of pain: it’s an XML-based standard tied to IFC models, and writing one means knowing both the grammar and the building-world rules behind it. ONESTRUCTION wanted the model to let people who are not BIM specialists review and manage attribute information anyway.
To get there, the team built Ishigaki-IDS on Qwen3 in 8B, 14B, and 32B sizes. The training pipeline had three stages. Continued pre-training injected IDS and IFC knowledge using web corpora and synthetic data created with internal domain experts. Supervised fine-tuning then taught the model from IDS authoring instructions in CSV or natural language paired with expected outputs. A third stage, reinforcement learning with verifiable rewards, used buildingSMART’s IDS-Audit-Tool so the model could be pushed by checks for XML well-formedness, IDS structure, and semantic consistency.
The infrastructure was straightforward and serious: two Amazon EC2 P5en.48xlarge nodes with NVIDIA H200 Tensor Core GPUs, AWS ParallelCluster, and Amazon FSx for Lustre for training data, synthetic data, and checkpoints. ONESTRUCTION and GenAIIC met every two weeks to review training results, data design, evaluation, and failures. They also built their own benchmark, IDS-Bench, covering IFC version, discipline, language, and implement, structure, and content axes.
On that benchmark, Ishigaki-IDS reportedly came close to 100 percent on XML structural compliance and IDS structural compliance, and rose above 80 percent on IDS content consistency. General frontier models could produce well-formed XML, but did far worse on the IDS-specific parts. The team also says the model handled inputs and outputs of roughly 120k tokens with YaRN, and that a proof-of-concept with buildingSMART drew positive reactions from both IDS specialists and non-specialists.
My take — AI-written commentary, not fact-checked reporting
This is the rare AI project that sounds boring until it works, which is usually how the useful ones behave. Construction does not need another chatbot with confidence issues; it needs models that can obey a standard and survive a checker's glare. The bigger lesson is pretty plain: in narrow fields, domain experts plus verifiable rewards beat grand speeches about general intelligence, every time.
Read more about this at: Amazon Web Services