ONESTRUCTION's Ishigaki-IDS brings AI to construction BIM via AWS pipeline

ONESTRUCTION, a construction technology startup, partnered with AWS GenAIIC under the GENIAC program to build Ishigaki-IDS, a domain-specialized foundation model for BIM (Building Information Modeling) workflows. The challenge: IDS (Information Delivery Specifications) is a new 2024 XML standard that defines what data must be attached to BIM models, and authoring it requires deep expertise in both IFC vocabulary and IDS grammar. Public training data is nearly nonexistent, and even web corpora lack depth. To solve this, ONESTRUCTION generated synthetic IDS files at scale with internal domain experts, then injected IFC vocabulary—mapping terms like "beam" to IfcBeam—and taught the model IDS-specific XML structure through a three-stage training pipeline: continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning with verifiable rewards (RLVR).

The model is built on Alibaba Cloud's open-source Qwen3 family (8B, 14B, and 32B parameters), chosen for strong multilingual support and flexible scaling. The team validated at smaller sizes before running full 32B training on Amazon EC2 P5en instances orchestrated with AWS ParallelCluster for stable distributed training. The RLVR stage uses verifiable rewards to enforce valid XML and correct IFC mappings, a key trick for structured output generation in low-resource domains. The result is practical: practitioners who aren't BIM specialists can now review and manage attribute information without manually authoring IDS files. For ML engineers, the reusable pattern—synthetic data, multi-stage training, and verifiable rewards—offers a blueprint for specializing foundation models where data is scarce, whether in construction, finance, healthcare, or other niche verticals.