How ONESTRUCTION built the Ishigaki-IDS foundation model with AWS GenAIIC

The construction technology startup ONESTRUCTION, Inc., in collaboration with Amazon Web Services Japan G.K., developed Ishigaki-IDS, a foundation model specialized for construction industry BIM workflows. This project was part of AWS’s Generative AI Accelerator Challenge Phase 3, with technical advisory from the AWS Generative AI Innovation Center (GenAIIC).

Building an IDS foundation model presented three main challenges: data scarcity, integrating an IFC vocabulary of several thousand terms, and IDS-specific grammar. To overcome these obstacles, ONESTRUCTION employed a three-stage training pipeline, collaborated closely with domain experts, and utilized infrastructure designed for stable distributed training.

The first stage, continued pre-training (CPT), involved injecting IDS and IFC domain knowledge using web corpora and synthetic data generated by internal domain experts. This synthetic data covered most of the training corpus and helped the model develop contextual understanding of IDS and related topics. The second stage, supervised fine-tuning (SFT), trained the model on pairs of IDS authoring instructions and their expected IDS output.

However, SFT alone did not resolve all issues, such as plausible but incorrect XML tag choices and wrong attribute values.

To address these remaining challenges, the third stage, reinforcement learning with verifiable rewards (RLVR), was introduced. RLVR utilized the IDS-Audit-Tool from buildingSMART International, a standards body, as the reward function. This tool verified XML well-formedness, IDS structural validity, and semantic consistency, allowing the model to iterate based on mechanical correctness signals.

Through this three-stage training pipeline, coupled with technical expertise from GenAIIC and Amazon Web Services Japan G.K., ONESTRUCTION successfully built Ishigaki-IDS, demonstrating how to overcome data scarcity and build specialized AI models in data-poor domains.