This submit was co-written by ONESTRUCTION, Inc. and Amazon Internet Companies Japan G.Ok. as a part of GENIAC (Generative AI Accelerator Problem) Part 3, with technical advisory from the AWS Generative AI Innovation Middle (GenAIIC).
Constructing domain-specialized basis fashions in data-scarce fields is difficult. You want sufficient coaching knowledge, specialised information, and methods to confirm your outputs.
ONESTRUCTION, Inc. is a building expertise startup that solves trade issues by way of openBIM. With technical advisory from GenAIIC, the corporate constructed Ishigaki-IDS, a basis mannequin (FM) specialised for building trade BIM (Constructing Data Modeling) workflows. BIM is a digital illustration of a constructing’s bodily and useful traits, used throughout the development lifecycle.
Japan’s building sector faces a persistent labor scarcity. BIM is promoted on the nationwide stage as a result of it lets design, building, and upkeep groups share data in a single place. However adopting BIM requires specialist information, and that studying value has slowed wider use. instance is IDS (Data Supply Specs), an XML-based normal that defines the data connected to and validated towards a BIM mannequin (an IFC (Trade Basis Lessons) mannequin). Authoring an IDS file takes fluency in its grammar plus information of IFC and its guidelines. Ishigaki-IDS lowers that barrier so practitioners who aren’t BIM specialists can overview and handle attribute data.
This submit is an architectural case research of how ONESTRUCTION constructed Ishigaki-IDS. For those who’re a machine studying (ML) engineer engaged on area adaptation, or a technical chief weighing methods to construct specialised AI fashions the place knowledge is scarce, you’ll discover a sample you possibly can reuse. Development and BIM professionals may even see what AI can do of their subject. Familiarity with basis mannequin coaching (pre-training and fine-tuning) and fundamental AWS compute ideas helps, nevertheless it isn’t required.
You’ll study:
Three challenges in constructing an IDS basis mannequin
Three issues stood between us and a working IDS mannequin.
The primary was knowledge shortage. IDS is a comparatively new normal, revealed in 2024, and building typically is a website with restricted public internet content material. Many different domains resembling finance, healthcare, and legislation prepare fashions on corpora of billions to a whole lot of billions of tokens, however no comparable public dataset exists for IDS. Even after accumulating latest internet knowledge, the amount was small and the depth was shallow, which meant the mannequin couldn’t decide up sufficient context about IDS and associated subjects from knowledge alone.
The second was injecting an IFC vocabulary of a number of thousand phrases. For instance, “beam” maps to IfcBeam and “air conditioner” maps to IfcUnitaryEquipment. This mapping has traditionally been achieved by hand by area specialists, and we would have liked the mannequin to study it straight.
The third was IDS-specific grammar. IDS is greater than plain XML: its tag construction modifications relying on what data is being connected or validated, and authors should use repeated patterns and devoted tags. Common-purpose basis fashions wrestle to supply this construction with accuracy.
Answer
Our method mixed three elements: a multi-stage coaching pipeline, shut collaboration with area specialists, and infrastructure constructed for steady distributed coaching. We begin with the coaching pipeline.
Three-stage coaching pipeline
We constructed Ishigaki-IDS on prime of Qwen3 (8B / 14B / 32B), an open-source massive language mannequin (LLM) from Alibaba Cloud recognized for sturdy multilingual capabilities and a variety of parameter sizes. With the dimensions vary, we will experiment at smaller scales earlier than committing to full coaching runs at 32B. We utilized a three-stage coaching pipeline.
First, in continued pre-training (CPT), we injected IDS and IFC area information utilizing internet corpora plus artificial knowledge created with our inside area specialists. We generated legitimate IDS information at scale and constructed artificial datasets that defined IDS-related paperwork from a number of angles, with artificial knowledge protecting a lot of the coaching corpus.
Second, in supervised fine-tuning (SFT), we skilled the mannequin on pairs of IDS authoring directions (in CSV or pure language) and their anticipated IDS output. SFT alone left anticipated points, resembling believable however incorrect XML tag decisions and incorrect attribute values, so we designed a 3rd stage to handle them.
Third, in reinforcement studying with verifiable rewards (RLVR), we used IDS-Audit-Instrument from buildingSMART, the worldwide requirements physique, because the reward perform. The instrument checks XML well-formedness, IDS structural validity, and semantic consistency, so the mannequin can iterate towards mechanical correctness alerts. RLVR matches the IDS activity nicely as a result of it refines output high quality with out massive quantities of supervised knowledge—helpful for a data-poor area.
Technical advisory from GenAIIC
We led growth with our building and BIM area experience and met with GenAIIC each two weeks for technical advisory. At every milestone, we introduced coaching outcomes and analysis knowledge to those periods, and collectively we labored by way of 5 key areas:
- Coaching knowledge design – artificial knowledge methods for the IDS area and methods to steadiness the information combine throughout CPT, SFT, and RLVR phases.
- Analysis benchmarks – metrics protecting IFC and IDS information, structured era, and normal dialogue capability.
- Coaching phases and methods – refining CPT, SFT, and RLVR, together with long-context dealing with, reward shaping, and structured era.
- Coaching infrastructure – parallelization, throughput, and stability for distributed coaching.
- Consequence analysis – when points appeared, diagnosing root causes and setting course for the subsequent iteration.
Iterating on “what change improves IDS era accuracy and practicality” at every cycle helped us construct a domain-specialized basis mannequin in a distinct segment, data-poor space inside a brief timeline.
Structure
For the coaching infrastructure, we used Amazon EC2 P5en situations (two p5en.48xlarge nodes with NVIDIA H200 Tensor Core GPUs), orchestrated with AWS ParallelCluster. ParallelCluster is an open supply instrument that simplifies deploying and managing Excessive Efficiency Computing (HPC) clusters on AWS. We saved coaching knowledge, artificial knowledge, and checkpoints on Amazon FSx for Lustre, a completely managed file system optimized for compute-intensive workloads that delivers sub-millisecond latencies and excessive throughput. This setup gave us steady multi-node distributed coaching and parallel entry to massive datasets.
Determine 1: Ishigaki-IDS coaching structure utilizing Amazon EC2 P5en situations orchestrated by AWS ParallelCluster with Amazon FSx for Lustre for high-throughput knowledge entry
Analysis
We constructed our personal analysis benchmark, IDS-Bench, with our inside IDS specialists. IDS-Bench measures efficiency throughout IFC model, building self-discipline (structure, construction, MEP, and customary), language (Japanese and English), and the Implement, Construction, and Content material axes, so the scores mirror what the mannequin must deal with in actual work.
Outcomes
In our IDS-Bench analysis, Ishigaki-IDS scored near one hundred pc on XML structural compliance and IDS structural compliance, and above 80 p.c on IDS content material consistency. Common frontier fashions instructed a unique story: they produced well-formed XML however scored below roughly 25 p.c on IDS structural compliance and close to 0 p.c on IDS content material consistency. IDS is a specialised and comparatively new space, which is the sort of downside a domain-specialized mannequin can remedy. The mannequin additionally helps context-length scaling with YaRN (Yet one more RoPE extensioN). YaRN extends the context window of transformer fashions past their authentic coaching size with out main efficiency degradation. We confirmed that the mannequin generates accurately with inputs and outputs as much as roughly 120k tokens.
In a joint proof-of-concept with buildingSMART, each IDS specialists and non-specialists responded positively to utilizing the mannequin of their work and to its capability to supply the meant IDS even from ambiguous prompts. In addition they gave us a listing of recommendations for additional growth, which strengthened our view that the mannequin is beneficial in follow.
Determine 2: IDS-Bench scores evaluating Ishigaki-IDS variants towards normal frontier fashions throughout XML construction, IDS construction, and IDS content material consistency
Classes realized
Three takeaways from this undertaking:
- Artificial knowledge high quality issues greater than amount. Our area specialists’ involvement in artificial knowledge creation was the difference-maker for mannequin efficiency. Quantity alone wouldn’t have produced the identical end result.
- Verifiable rewards speed up iteration. Utilizing
IDS-Audit-Instrumentas an automatic reward sign allow us to iterate sooner than guide analysis would permit, particularly in a data-poor setting. - Steady infrastructure lets us experiment freely. Dependable distributed coaching on Amazon EC2 P5en, AWS ParallelCluster, and Amazon FSx for Lustre freed us to give attention to mannequin enhancements reasonably than debugging cluster points.
Conclusion
Combining area skilled collaboration, artificial knowledge, and RLVR tied to a verification instrument labored nicely for constructing a domain-specialized mannequin in a data-poor specialty space. Steady technical advisory from GenAIIC helped us attain the accuracy targets measured on IDS-Bench inside the GENIAC Part 3 timeline. ONESTRUCTION will proceed working with AWS to convey AI instruments to the development trade.
Subsequent steps
For those who’re all for constructing domain-specialized basis fashions to your trade, the next sources are place to start out:
Concerning the authors
