Textual description
A deterministic, template-based natural language description generated automatically for each model, covering classes, attributes, enumerations, generalizations, and associations.
Results · Dataset
The B-UML Dataset
A large-scale, multi-representation dataset of 9,081 structurally validated B-UML models spanning 302 categories, designed to support empirical software engineering research and AI-based modeling tasks such as LLM fine-tuning and benchmarking.
- 9,081
- Validated models
- 302
- Domain categories
- 564,812
- Modeling elements
- 3
- Source datasets
What's inside
Five artifacts per model
Every entry in the dataset goes beyond a raw model file. Each of the 9,081 B-UML models is accompanied by a complete set of verified, automatically generated artifacts.
Visual representation
A UML-compliant class diagram image generated via a custom Graphviz-based renderer, providing an immediate visual understanding of each model.
Executable code
Compilable and executable code in both Python and Java, automatically generated from B-UML models, including attributes, associations, and cardinality constraints.
Category and metadata
Each model is labelled with a domain category (302 in total) and structural metadata such as the number of classes, attributes, and associations.
Test cases
Automatically generated test cases for each model, ensuring alignment between the conceptual specification and the produced outputs.
Provenance
Dataset sources
The dataset combines three publicly available MDE datasets with real-world models from applications built on the BESSER platform.
ModelSet
5,466 Ecore metamodels from GitHub and 5,120 UML models from GenMyModel. Only class diagrams are included.
Ecore-555
555 Ecore metamodels mined from GitHub, each manually labelled across nine domain categories.
Pan et al.
52 unique metamodels with 369 OCL constraints alongside their natural language specifications.
Built with BESSER
Real-world applications built with BESSER: NexaCRM, AI Sandbox, and Skillset Matcher.
Quality assurance
Three-level validation
All models pass a rigorous automated validation pipeline before inclusion. Of 11,196 candidate models, 9,081 were retained after removing structurally invalid entries.
- 01
Inheritance checks
Models with the diamond inheritance problem, where a class inherits the same superclass through two or more paths, are removed to ensure structural soundness.
- 02
Model-level verification
Each B-UML model is validated for loadability in the editor and element-count consistency between the source and the converted B-UML representation.
- 03
Code-level verification
Generated Python and Java code is checked for both compilability (no syntax errors) and executability (no runtime errors on instantiation).
Explore the data
Browse the dataset interactively
Use the interactive web explorer to browse, filter, and inspect individual B-UML models, their generated artifacts, and category distribution across the full dataset.