A construction model tests domain AI where data is scarce
ONESTRUCTION says it built Ishigaki-IDS, a foundation model specialized for construction BIM workflows, with technical advisory from the AWS Generative AI Innovation Center. The project targets Information Delivery Specifications, an XML-based standard used to define and validate information attached to BIM models. The case study is notable because it shows a domain-AI pattern based on synthetic data, expert review and verifiable rewards rather than broad web-scale data alone.Why BIM-specific AI is difficult
Ishigaki-IDS addresses a narrow but practical construction problem: creating and checking IDS files for BIM models. According to the source, IDS files require knowledge of IDS grammar, IFC rules and domain-specific attribute requirements, which can make adoption difficult for practitioners who are not BIM specialists.The challenge is amplified by limited public training material. ONESTRUCTION describes IDS as a relatively new standard, published in 2024, and says there is no comparable public dataset for IDS like the very large corpora available in better-documented fields. The model therefore had to learn from a scarce domain rather than from abundant public text.
The work also required learning an IFC vocabulary of several thousand terms. The source gives examples such as mapping beam to IfcBeam and air conditioner to IfcUnitaryEquipment. If that mapping is wrong, a generated IDS may look plausible while failing to express the required construction object or attribute.
The training stack behind Ishigaki-IDS
ONESTRUCTION built Ishigaki-IDS on Qwen3, using 8B, 14B and 32B parameter variants, and applied a three-stage training pipeline. The stages were continued pre-training, supervised fine-tuning and reinforcement learning with verifiable rewards.In continued pre-training, the company used web corpora and synthetic data created with internal domain experts to inject IDS and IFC knowledge. The case study says synthetic data covered most of the training corpus, including valid IDS files and explanatory datasets that represented IDS-related documents from multiple angles.
Supervised fine-tuning then trained the model on instruction-output pairs. Inputs could be CSV or natural-language authoring requests, while outputs were expected IDS files. ONESTRUCTION says SFT alone still produced problems such as plausible but incorrect XML tags and wrong attribute values, which led to the third training stage.
Verifiable rewards narrowed the output problem
The reinforcement stage used the IDS-Audit-Tool from buildingSMART as the reward function. That tool checks XML well-formedness, IDS structural validity and semantic consistency, giving the model a mechanical signal for whether a generated file satisfies the standard.This matters because IDS generation is not only a language task. A response can be fluent and still fail if its XML structure, IDS tags or IFC-related content are wrong. By using verifiable rewards, ONESTRUCTION could refine outputs without relying only on manual review or large supervised datasets.
The approach is useful for other data-poor technical domains only if there is a reliable validation mechanism. In this case, the reward function matched the task closely because the output is structured and can be checked against explicit rules.
AWS infrastructure supported multi-node training
The source says training ran on Amazon EC2 P5en infrastructure, using two p5en.48xlarge nodes with NVIDIA H200 Tensor Core GPUs. AWS ParallelCluster orchestrated the cluster, while Amazon FSx for Lustre stored training data, synthetic data and checkpoints for high-throughput access.ONESTRUCTION frames the setup as a way to keep distributed training stable enough for repeated experiments. That point is practical: in a specialized model project, the value often comes from iterating on data mix, reward design and evaluation rather than from a single training run.
The technical advisory relationship with AWS GenAIIC was also structured around iteration. The source says ONESTRUCTION met with GenAIIC every two weeks to review training results, evaluation data, infrastructure behavior, benchmark design and diagnosis of model issues.
Benchmark results show the domain gap
ONESTRUCTION built an internal benchmark called IDS-Bench with its IDS specialists. The benchmark measures IFC version, construction discipline, Japanese and English language handling, and three axes described as Implement, Structure and Content.According to the case study, Ishigaki-IDS scored close to 100 percent on XML structural compliance and IDS structural compliance, and above 80 percent on IDS content consistency. The same source says general frontier models produced well-formed XML but scored under roughly 25 percent on IDS structural compliance and near 0 percent on IDS content consistency.
Those figures should be read as reported benchmark results from the project, not as an independent audit. Still, they illustrate the central technical claim: a general model may handle surface syntax while missing domain structure, whereas a specialized model can be tuned for the rules that matter in production workflows.
Proof-of-concept feedback and remaining limits
The model also supports context-length scaling with YaRN, which ONESTRUCTION says allowed correct generation with inputs and outputs up to roughly 120k tokens. Long context can matter in BIM workflows because relevant instructions and model information may be spread across large documents.In a joint proof-of-concept with buildingSMART, the source says IDS specialists and non-specialists responded positively to using the model and to its ability to produce intended IDS outputs from ambiguous prompts. Participants also provided suggestions for further development, which indicates the system is not being presented as finished infrastructure.
The broader implication is measured but significant. If a domain has formal outputs, expert-curated synthetic data and a validator that can act as a reward signal, specialized AI may be easier to evaluate than open-ended assistants. Construction BIM is one example, not a guarantee for every professional field.
Conclusion
Ishigaki-IDS is a focused example of domain adaptation in a field where public data is limited and correctness is rule-bound. ONESTRUCTION combined Qwen3, expert synthetic data, supervised training, verifiable rewards and AWS distributed training infrastructure to target IDS generation for BIM workflows.The most useful lesson is not that every construction task needs its own foundation model. It is that specialized AI projects need a tight connection between domain knowledge, measurable output quality and infrastructure that supports iteration. For BIM validation, ONESTRUCTION says that combination produced stronger IDS-specific results than general frontier models in its internal benchmark.
Sources
Editorial Team - CoinBotLab
