ONESTRUCTION develops Ishigaki-IDS foundation model for construction BIM workflows with AWS GenAIIC

Ishigaki-IDS Foundation Model Development

ONESTRUCTION, Inc., a construction technology startup, developed Ishigaki-IDS, a foundation model (FM) tailored for Building Information Modeling (BIM) workflows in the construction industry. This development was supported by technical advisory from the AWS Generative AI Innovation Center (GenAIIC) as part of the GENIAC (Generative AI Accelerator Challenge) Phase 3.

Addressing Data Scarcity in Specialized Domains

A key challenge in building domain-specialized foundation models, particularly in fields with limited data, is the scarcity of training data. Ishigaki-IDS addresses this by utilizing synthetic data generation. The model employs a three-stage training pipeline: CPT (Continued Pre-training), SFT (Supervised Fine-tuning), and RLVR (Reinforcement Learning from Verifiable Rewards).

Simplifying BIM Workflows

BIM is a digital representation of a building's characteristics used throughout its lifecycle. Japan's construction sector promotes BIM to combat labor shortages, but its adoption is hindered by the specialized knowledge required. Ishigaki-IDS aims to lower this barrier, specifically for Information Delivery Specifications (IDS), an XML-based standard for validating information against a BIM model. The model allows practitioners without BIM specialization to review and manage attribute information.

Architectural Case Study

The development of Ishigaki-IDS serves as an architectural case study for machine learning engineers working on domain adaptation and technical leaders considering specialized AI models in data-scarce environments. It demonstrates how to build a specialized AI model using synthetic data, a multi-stage training pipeline, and verifiable rewards for structured output generation. Distributed training was run on Amazon EC2 P5en instances with AWS ParallelCluster.