ONESTRUCTION, Inc., a construction technology startup, has developed the Ishigaki-IDS foundation model, a specialized AI tool for construction industry BIM workflows. The model was built with technical advisory from the AWS Generative AI Innovation Center (GenAIIC) as part of the GENIAC (Generative AI Accelerator Challenge) Phase 3. The project addresses the challenge of creating domain-specific foundation models in data-scarce fields by leveraging synthetic data generation and a three-stage training pipeline.
The Ishigaki-IDS model was built on top of Qwen3, an open-source large language model from Alibaba Cloud, and uses a three-stage training pipeline. The first stage, continued pre-training (CPT), involved injecting IDS and IFC domain knowledge using web corpora and synthetic data. The second stage, supervised fine-tuning (SFT), trained the model on pairs of IDS authoring instructions and expected outputs. A third stage, reinforcement learning with verifiable rewards (RLVR), used the IDS-Audit-Tool from buildingSMART to refine output quality. This approach enabled the model to generate accurate XML tags and attribute values for BIM workflows.
The project faced three main challenges: data scarcity, IFC vocabulary injection, and IDS-specific grammar. These were addressed through synthetic data generation, collaboration with domain experts, and infrastructure built for stable distributed training. The model was trained on Amazon EC2 P5en instances using AWS ParallelCluster and stored data on Amazon FSx for Lustre. The result is a domain-specialized foundation model that performs well on structured XML and IDS compliance tasks.
Source: awsml
