The construction tech startup used synthetic data and a three-stage pipeline to specialize a model for IDS standards.

ONESTRUCTION, a construction technology startup, built Ishigaki-IDS, a foundation model specialized for BIM workflows, with technical advisory from AWS GenAIIC. The model targets the XML-based IDS standard, which defines what information a BIM model must include and how to validate it. Ishigaki-IDS lowers the barrier so practitioners who are not BIM specialists can review and manage attribute information.

The project runs under GENIAC Phase 3, a Japanese accelerator for generative AI, with AWS GenAIIC offering technical advisory. Japan promotes BIM to ease a construction labor shortage, but specialist knowledge has slowed adoption. IDS debuted in 2024, and public data for the standard is scarce, so the team needed synthetic data.

The three-stage pipeline combines continued pretraining, supervised fine-tuning, and reinforcement learning with verifiable rewards. Verifiable rewards keep structured output like IDS files accurate, and distributed training runs on EC2 P5en instances with AWS ParallelCluster. ML teams in any data-scarce domain can reuse this pattern, and builders can expect lower BIM adoption barriers.

The Ishigaki-IDS project shows how synthetic data and verifiable rewards can make domain-specific foundation models practical. Other teams facing scarce data can follow the same three-stage recipe. Watch for more specialized models in industries where standards are new or data is thin. ONESTRUCTION’s work also points to deeper AI integration in construction and openBIM workflows.

What matters

• ONESTRUCTION built Ishigaki-IDS, a BIM-specialized foundation model with AWS GenAIIC.
• Ishigaki-IDS lets non-BIM specialists review and manage attribute information without deep IFC expertise.
• The three-stage training pipeline of CPT, SFT, and RLVR offers a reusable pattern for data-scarce domains.

Why it matters

The three-stage training pipeline of CPT, SFT, and RLVR offers a reusable pattern for data-scarce domains.

This GenAI News article was prepared in original wording using reporting and materials published by AWS Machine Learning Blog. Source reference: https://aws.amazon.com/blogs/machine-learning/how-onestruction-built-the-ishigaki-ids-foundation-model-with-aws-genaiic/.

Drafted by the GenAI News review pipeline.