ONESTRUCTION Develops Construction Domain Model with AWS Generative AIIC Support

ONESTRUCTION, a construction technology startup, collaborated with AWS's Generative AI Innovation Center (GenAIIC) to develop Ishigaki-IDS, a foundation model designed specifically for construction industry BIM (Building Information Modeling) workflows. This project was part of the GENIAC Phase 3 challenge, aiming to address the challenges of building domain-specific AI models in data-scarce fields.

The construction sector in Japan faces a labor shortage, and BIM is promoted to facilitate information sharing among design, construction, and maintenance teams. However, the adoption of BIM is hindered by the need for specialized knowledge, particularly in authoring IDS (Information Delivery Specifications) files, which are XML-based standards requiring fluency in complex grammar and IFC vocabulary.

To overcome these challenges, ONESTRUCTION employed a three-stage training pipeline on top of Alibaba Cloud's open-source large language model, Qwen3. The pipeline included continued pre-training with synthetic data, supervised fine-tuning with instruction-output pairs, and reinforcement learning with verifiable rewards using the IDS-Audit-Tool from buildingSMART. This approach allowed the model to learn domain-specific vocabulary, grammar, and structural compliance.

The training infrastructure utilized Amazon EC2 P5en instances orchestrated with AWS ParallelCluster, with data stored on Amazon FSx for Lustre, enabling stable distributed training. The resulting model, Ishigaki-IDS, was evaluated using a custom benchmark, IDS-Bench, measuring performance across various dimensions relevant to construction workflows.

Results showed that Ishigaki-IDS achieved near-perfect scores in XML and IDS structural compliance and over 80% in content consistency, significantly outperforming general-purpose models. The model also demonstrated the ability to generate accurate IDS files with inputs up to approximately 120,000 tokens, supporting long-context scenarios.

The project yielded key insights: the importance of synthetic data quality over quantity, the effectiveness of verifiable rewards in accelerating iteration, and the necessity of stable infrastructure for experimentation. The collaboration with domain experts and continuous technical advisory from GenAIIC were instrumental in achieving these outcomes.

Looking ahead, ONESTRUCTION plans to further refine the model and expand its capabilities, with resources available for others interested in developing domain-specific AI models using AWS infrastructure and tools. The Ishigaki-IDS model is accessible on Hugging Face for testing and deployment in real-world scenarios.