Construction startup builds BIM-specialized AI model with AWS GenAIIC

ONESTRUCTION, Inc., a construction technology startup, set out to solve a real industry problem: BIM adoption in Japan's construction sector is promoted at the national level because it lets design, construction, and maintenance teams share information in one place, but adoption has been slowed by the steep learning cost of specialist tools. One such tool is IDS (Information Delivery Specifications), an XML-based standard that defines the information attached to and validated against a BIM model (an IFC model). Authoring an IDS file historically requires fluency in its grammar plus deep knowledge of IFC and its rules. ONESTRUCTION built Ishigaki-IDS, a foundation model specialized for BIM workflows, to lower that barrier so practitioners who aren't BIM specialists can review and manage attribute information.

The company faced three core challenges. First, IDS is a relatively new standard (published in 2024) and construction is a domain with limited public web content, so no comparable public dataset exists for training—unlike finance, healthcare, or law, which train on billions to hundreds of billions of tokens. Second, the model needed to learn an IFC vocabulary of several thousand terms (mapping "beam" to IfcBeam, "air conditioner" to IfcUnitaryEquipment, and so on), work that has historically been done by hand. Third, IDS has specialized grammar: its tag structure changes depending on what information is being attached or validated, and authors must use repeated patterns and dedicated tags—general-purpose models struggle to produce this structure with accuracy.

ONESTRUCTION's solution combined three ingredients: a multi-stage training pipeline, close collaboration with domain experts, and stable distributed training infrastructure. The team built Ishigaki-IDS on top of Qwen3 (8B / 14B / 32B), an open-source large language model from Alibaba Cloud known for strong multilingual capabilities. In continued pre-training (CPT), they injected IDS and IFC domain knowledge using web corpora plus synthetic data created with internal domain experts; synthetic data covered most of the training corpus. In supervised fine-tuning (SFT), they trained the model on pairs of IDS authoring instructions (in CSV or natural language) and their expected IDS output. In reinforcement learning with verifiable rewards (RLVR), they used IDS-Audit-Tool from buildingSMART as the reward function, so the model could iterate against mechanical correctness signals—checking XML well-formedness, IDS structural validity, and semantic consistency.

For infrastructure, ONESTRUCTION used Amazon EC2 P5en instances (two p5en.48xlarge nodes with NVIDIA H200 Tensor Core GPUs), orchestrated with AWS ParallelCluster, with training data and checkpoints stored on Amazon FSx for Lustre. ONESTRUCTION met with the AWS Generative AI Innovation Center (GenAIIC) every two weeks for technical advisory, working through training data design, evaluation benchmarks, training stages and techniques, training infrastructure, and result diagnosis at each milestone.

On ONESTRUCTION's own evaluation benchmark, IDS-Bench, Ishigaki-IDS scored close to 100 percent on XML structural compliance and IDS structural compliance, and above 80 percent on IDS content consistency. General frontier models produced well-formed XML but scored under roughly 25 percent on IDS structural compliance and near 0 percent on IDS content consistency. The model also supports context-length scaling with YaRN (Yet another RoPE extensioN), generating correctly with inputs and outputs up to roughly 120k tokens. In a joint proof-of-concept with buildingSMART, both IDS specialists and non-specialists responded positively to using the model in their work and to its ability to produce the intended IDS even from ambiguous prompts. ONESTRUCTION identified three key lessons: synthetic data quality matters more than quantity, verifiable rewards accelerate iteration in data-poor settings, and stable infrastructure lets teams experiment freely rather than debugging cluster issues.