ardalon/semaadb-dataset

SEMAADB

SEMAADB is a dataset of 3,000 engineering contexts and 15,000 connected SysML diagrams. Each context contains Requirement, Block Definition, Activity, State Machine, and Sequence views of the same system. The diagrams are represented as PlantUML source and rendered SVG.

Dataset structure

The train split contains 2,900 automatically checked contexts. The test split contains the frozen 100-context human-verified benchmark core. A context is the unit of data: its system description, shared entity model, five diagrams, verification tier, and generation telemetry remain together in one row. Public IDs run from ctx-0001 through ctx-3000.

Each named entry under diagrams contains plantuml, svg, render-attempt counts, and any automatic consistency warnings. generation records the timestamp, input/output token counts for the description and modeling stages, and wall-clock duration.

To inspect the format without downloading either complete split, browse the sample/ directory. It contains five representative context folders. Each keeps context.json at its root and places the five PlantUML sources under code/ and their rendered SVGs under render/.

Loading and querying

from datasets import load_dataset

dataset = load_dataset("ardalon/semaadb-dataset", revision="v1.0.0")
print(dataset)

Find a specific context:

context = dataset["test"].filter(
    lambda row: row["context_id"] == "ctx-0004"
)[0]
print(context["system_description"])
print(context["diagrams"]["requirement"]["plantuml"])

Query a system type, domain, or scale:

gearboxes = dataset["train"].filter(
    lambda row: row["system_type"] == "gearbox / transmission"
    and row["scale"] == "system"
)
embedded = dataset["train"].filter(
    lambda row: row["domain"] == "cs-embedded"
)

Display a rendered diagram in a notebook:

from IPython.display import SVG, display

display(SVG(data=context["diagrams"]["state"]["svg"]))

Download every release file without loading the table:

from huggingface_hub import snapshot_download

release_dir = snapshot_download(
    repo_id="ardalon/semaadb-dataset", repo_type="dataset", revision="v1.0.0"
)

Dataset creation

The contexts were synthesized from public engineering references and terminology, with source URLs retained for grounding and traceability. OpenAI GPT-5.5 generated the system descriptions, shared entity models, and diagram sources. Every released diagram rendered successfully with PlantUML 1.2026.6. The generation prompts and validation methodology are described in the paper.

Intended uses

The dataset supports research and development in diagram generation, diagram understanding, semantic repair, cross-view consistency, and AI-assisted systems engineering. Keep the test split isolated when comparing systems against the paper's benchmark.

It is not a substitute for engineering approval, safety analysis, certification evidence, or an authoritative SysML model. Generated requirements and parameters must not be used directly in a real design without qualified review.

Limitations and responsible use

Only the 100-context test split was reviewed by people. auto-passed means that structural, rendering, content, and consistency checks completed; it does not claim human semantic verification. Automatic consistency warnings are retained in the records. The corpus is synthetic, uses five SysML v1 views encoded through PlantUML conventions, and may contain domain errors, omissions, or modeling choices with several reasonable alternatives.

Generation timestamps, token counts, and wall-clock values describe the original run. They are not billing records and should not be treated as portable performance benchmarks.

License and citation

SEMAADB v1.0.0 is released under CC BY 4.0. Grounding URLs identify references used during synthesis; linked third-party material remains under its own terms and is not redistributed here.

@misc{aryashad2026semaadb,
  title={A Validated Dataset and Benchmark for Coherent Multi-Diagram SysML Models},
  author={Ardalan Aryashad and Yan Jin},
  year={2026},
  url={https://huggingface.co/datasets/ardalon/semaadb-dataset}
}

Maintenance and versioning

This card describes immutable release v1.0.0. Corrections or additions receive a new tagged version and a documented change log. Checksums and release metadata are included in the repository.