Datasets:

OAEI-ML
/

bio-ml

license: other
license_name: bioml-2026-mixed
pretty_name: BioML 2026 (OAEI Bio-ML)
viewer: false
tags:
  - ontology-matching
  - ontology-alignment
  - biomedical
  - knowledge-graphs
  - owl
  - OAEI
  - UMLS
  - SNOMED-CT
size_categories:
  - 10K<n<100K

BioML 2026 — OAEI Bio-ML

The public task data for the OAEI Bio-ML 2026 track (a larger LargeBio), over three whole-ontology pairs: NCIT–DOID, SNOMED–FMA, SNOMED–NCIT. This repository holds the redistributable pieces — the reference/candidate data as entity IRIs/CURIEs and the manifests. The ontologies themselves are not re-hosted (see Ontologies); the held-out test gold is organiser-side only; the scoring kit is code and lives in the GitHub repo (see Quickstart). Ssee the evaluation metrics on the track website.

Tasks

Global equivalence alignment — submit an OAEI Alignment RDF/XML or a TSV per pair. Scored (organiser-side) with repaired coherence-aware P/R/F1 (headline) + standard P/R/F1 + Global Coherence. A public refs_equiv/train.tsv of positive mappings is provided per pair.
Local equivalence ranking — rank the 100-candidate pool per query. Scored with MRR + Hits@{1,5,10}.

Layout (per pair `NCIT-DOID/`, `SNOMED-FMA/`, `SNOMED-NCIT/`)

Path	Contents
`local.{train,valid,test}.cands.tsv`	Subtrack 2 ranking pools (100 cands/query; test is gold-stripped)
`refs_equiv/train.tsv`	Subtrack 1 public train positives (`SrcEntity, TgtEntity, Score`)
`repaired/`	the same set against the coherence-repaired reference
`evaluation/`	`sample_submission.tsv`, `submission_schema.json` (shared)
`SHA256SUMS.txt`, `release_manifest.json`, `license_manifest.json`, `resolved_versions.json`	integrity + provenance

The scoring kit (validators + self-scorers, zero-install) is code and lives in the GitHub repository liseda-lab/OAEI-Bio-ML, not in this dataset.

Coming from the original Bio-ML?

local.*.cands.tsv corresponds to Bio-ML's refs_equiv/test.cands.tsv (now a 3-way train/valid/test split, 100 cands/query). refs_equiv/train.tsv is the supervised positive-mapping file. The ontologies are not bundled here.

Ontologies

The task uses the whole source ontologies; obtain them yourself (versions + checksums pinned in resolved_versions.json, instructions in ontologies.md): NCIT, MONDO, DOID are openly downloadable (CC BY / CC0); SNOMED CT and UMLS require your own UTS licence; FMA is downloadable from its source. See license_manifest.json.

Quickstart

The scoring kit lives in the GitHub repo — clone it for the tools, download this dataset for the data:

git clone https://github.com/liseda-lab/OAEI-Bio-ML          # provides scoring_kit/
pip install -U huggingface_hub
hf download OAEI-ML/bio-ml --repo-type dataset --revision 2026 --local-dir ./bio-ml   # the data
python3 OAEI-Bio-ML/scoring_kit/self_check.py --data ./bio-ml            # oracle -> perfect scores
python3 OAEI-Bio-ML/scoring_kit/validate_ranking.py bio-ml/NCIT-DOID/local.test.cands.tsv my_ranking.tsv
python3 OAEI-Bio-ML/scoring_kit/score_local.py my_ranking.tsv bio-ml/NCIT-DOID/local.valid.cands.tsv

Submit via the CodaBench competitions (linked from the track website).

Licensing

The packaging and reference/candidate IRI data are provided for this evaluation and ontology-matching research. Underlying sources retain their licences: NCIT, MONDO CC BY 4.0; DOID CC0 1.0; SNOMED CT (Affiliate) and UMLS (Metathesaurus) are restricted (UTS) — their derived OWL/gold is not redistributed here; NCIT, MONDO, DOID, FMA are freely redistributable (OWL is pointer-only here, for size). Per-source details: license_manifest.json. Do not attempt to reconstruct the held-out test reference.

Citation

Cite the OAEI Bio-ML 2026 track (details on the track website).