Data

Public graph package, hidden-label evaluation

Five ontologies — NCIT, DOID, ORDO, SNOMED, FMA — are projected into one knowledge graph. Public train (60%) and validation (10%) carry gold labels; the private test split (30%) releases candidates with gold withheld.

What's in the package

Public release files and their purpose
Path	Purpose
`graph/triples.csv`	Typed intra-ontology graph triples and the released training anchors.
`graph/properties.csv`	Node identifiers, ontology membership, labels, synonyms, definitions, and permitted metadata.
`graph/.dl, graph/.facts, graph/datalog_terms.tsv`	The OWL 2 RL Datalog bundle: `facts.dl` (a Soufflé `.input` driver) with one `<relation>.facts` TSV per fact relation, `owl2rl_core.dl` (rule core), `rules.dl` (closure driver), `conflict_rules.dl` (conflict driver), the `datalog_terms.tsv` symbol dictionary and `datalog_schema.json`.
`alignments/{train,valid}.tsv`	Public supervised labels with their relation types.
`tasks//.cands.tsv`	Candidate files for train, validation, and test queries.
`tasks/*/{train,valid}.preferred.tsv`	One preferred (target, relation) gold per query; drives the Preferred Typed (Relation-Aware) MRR.
`tasks/*/{train,valid}.graded.tsv`	Graded relevance over (candidate, relation) pairs, from the preferred pairs and the target-ontology hierarchy; drives Hierarchy-Aware Typed nDCG@10.
`evaluation/*`	Sample submission, submission schema, and the local scorer. Kit scoring available; CodaBench integration in preparation. Provisional
`baseline_predictions/*`	Organiser baseline-system predictions on the test split.
`data_card.md, *_manifest.json, reports/`	The data card, licence and release manifests, and dataset reports.

Released as CSV, TSV, Datalog, JSON. The kit's data card is the authoritative inventory and file-format reference.

Ontology licences

Licence for each track ontology
Ontology	Licence
NCIT	CC BY 4.0
DOID	CC0 1.0
ORDO	CC BY 4.0
FMA	FMA License (redistribution terms apply)
SNOMED	SNOMED CT Affiliate License (redistribution terms apply)

Hosting & distribution

The dataset is distributed as a standalone public artefact, planned as a Hugging Face dataset release with an offline archive on Zenodo; the DOI is not yet finalised. Provisional