Procedural generation
Generated from BIM rules, not from real project files.
Procedurally generated synthetic BIM buildings for AI research and model training
AI-ready synthetic BIM datasets
Buildata creates fully synthetic BIM datasets for graph learning, machine learning and BIM language model workflows. Each package includes structured building data, semantic elements, relationships and training-ready records.
three layers in one dataset
01
Core BIM graph
Semantic building elements, metadata and relationships.
02
ML layer
Training-ready task records for prediction and benchmarking.
03
LLM layer
Instruction and reasoning samples for BIM AI workflows.
1 sample
1 complete building
3 export layers
Graph, ML and LLM
AI-ready
Structured for training
about the data
Synthetic by design
All Buildata catalog datasets are synthetic. Buildings are generated procedurally using BIM semantic rules based on IFC schemas, materials, property sets and spatial relationships.
No real-building data
Designed for public sharing and AI experimentation.
Built for AI
Structured for graph learning, ML pipelines and BIM LLMs.
what makes buildata different
More than synthetic BIM files
Buildata packages each building as structured AI-ready data, so teams can work directly with semantic graphs, benchmark tasks and training samples instead of starting from raw files.
For Graph Neural Networks
Use nodes, edges and topology directly in graph-based learning workflows.
For machine learning
Use generated task records for classification, prediction and validation experiments.
For BIM LLMs
Train language models on BIM instructions, reasoning and question-answering samples.
dataset layers
What a Buildata package can include
Core export
Metadata, nodes, edges, tasks, statistics and dataset card.
AI-ready layers
Graph exports, ML-ready tasks and LLM-oriented training samples.
Per-building packaging
Individual building samples for incremental testing and evaluation.
featured datasets
Explore AI-ready BIM datasets
Example instruction sample
{
"instruction": "Identify the IFC entity from the BIM element properties.",
"input": "Height: 3.2 m\nThickness: 0.2 m\nMaterial: Concrete\nObject type: Exterior Wall\nLoad bearing: true",
"output": "IfcWall"
}
Compatible workflows
- PyTorch and TensorFlow pipelines
- PyTorch Geometric and DGL experiments
- Hugging Face instruction tuning
- BIM reasoning and QA workflows
start here
Download a free sample
Start with a compact sample dataset and review the packaging, statistics and dataset card before moving to larger collections.