This dataset contains 3,478 paired BPMN 2.0 XML files and PNG renderings used to study structured extraction from business-process diagrams. Visible labels in the released pairs are English translations. The corresponding preparation, training, prediction, evaluation, and statistical-analysis code is available at:

https://github.com/pritamdeka/bpmn-structured-extraction

Frozen splits

train: 2,782 rows
validation: 347 rows
test: 349 rows

The repository preserves the frozen 80/10/10 split used in the reported experiments. The original exercise or scenario identifiers were not retained before the files were flattened. An identifier-invariant graph audit therefore finds exact structural overlap across splits, including 75 test rows whose graph fingerprint occurs in training or validation. Users should not describe these splits as scenario-independent or leakage-controlled.

Columns

filename: stable base filename, for example process_0001
png: rendered process-diagram image
bpmn_code: complete BPMN 2.0 XML text, including BPMN-DI geometry

The GitHub preparation script converts bpmn_code deterministically into the canonical JSON target used for supervision and writes the full split audit.

Intended uses

The dataset supports image-to-structure extraction, BPMN diagram understanding, XML/graph reconstruction, prompt evaluation, and parameter-efficient adaptation of vision-language models. It is not intended as evidence of performance on fully novel business scenarios without a separate provenance-controlled test.

Reproducibility

Use the immutable dataset revision together with the model revision, prompt, parser version, seed, and evaluation configuration. Do not place Hugging Face tokens in scripts or derived datasets.

Downloads last month: 15

Total file size:

321 MB