ODP-Reuse Datasets

The ODP-Reuse dataset collection provides a curated suite of benchmark datasets for studying and evaluating the reuse of Ontology Design Patterns (ODPs) in ontologies. Each dataset in this collection corresponds to a distinct discovery source—(i) Paper & Repository Search, (ii) GitHub Code Search, or (iii) DOS–DP template-based ontologies—and contains ontologies that have been verified to fully reuse at least one ODP under a strict 100% IRI-match criterion. Corpus-hashed IRI-anonymised twin of [`corpus`](../corpus). Same content, same reuse relations -- only identifiers differ. Every IRI is replaced by `http://www.mydata.org/` + URL-safe base64 of its SHA-1, the same function `build_release.py` uses for the hashed views, so all parts of the release interoperate: an entity reused by two ontologies keeps one identifier here too. RDF, RDFS, OWL and XSD IRIs are left intact so the files stay valid OWL. Blank-node labels are hashed deterministically. Literals are untouched. Why does an anonymised twin exist? A detection method evaluated on the plain corpus can cheat: it can recognise `ontologydesignpatterns.org` namespace prefixes instead of learning structure. Here that shortcut is gone, so a score reflects structural and semantic signal. Two things to know before using it Serialisation. Every file is N-Triples, whatever the original was. The transform streams -- it never holds a whole graph -- because the OBO portion includes single files up to 1.1 GiB that do not fit in memory as an rdflib graph. Re-serialising RDF/XML would require exactly the graph being avoided. `format_map.csv` gives the original extension and byte size for every file. Paths are not hashed. Directory and file names still carry ontology ids, so this tree aligns with `../corpus`, `../MANIFEST.csv` and the view tables. The anonymisation covers identifiers *inside* the files. If an experiment must not see ontology identity either, read through `format_map.csv` and rename. Files: 6092 of 6358 rewritten, 539 triples, 266 could not be parsed (see `hash_failures.csv`).