OAEI Bio-ML 2026

Quickstart

Everything the scoring kit runs is Python 3.12+ standard library only. Download the data, validate your files, and self-score where you hold the gold.

This guide takes you from an empty directory to a validated (and, where possible, self-scored) submission. Everything the scoring kit runs is Python 3.12+ standard library only.

1. Get the scoring kit and the data

The scoring kit (validators + self-scorers) and the task data come from two places: the kit ships with the track’s GitHub repository, and the data is published on the Hugging Face Hub. Get both, side by side:

# 1a. the scoring kit — clone (or download) the track repository
git clone https://github.com/liseda-lab/OAEI-Bio-ML
cd OAEI-Bio-ML

# 1b. the task data — download the Hugging Face dataset into ./bio-ml
pip install -U huggingface_hub
hf download OAEI-ML/bio-ml --repo-type dataset --revision 2026 --local-dir ./bio-ml

The 2026 task data is publicly available under the OAEI-ML organisation — huggingface.co/datasets/OAEI-ML/bio-ml, edition tag 2026 — and downloads freely, without gating (entity IRIs only). Under bio-ml/, each pair (NCIT-DOID, SNOMED-FMA, SNOMED-NCIT) contains:

  • refs_equiv/train.tsv — the public equivalence training reference for global alignment (SrcEntity, TgtEntity, Score; full IRIs; semi-supervised setting),
  • local.train.cands.tsv / local.valid.cands.tsv — the local-ranking pools that carry the gold TgtEntity (use them to self-score), and local.test.cands.tsv — the gold-stripped test pool (source entity + candidate list only),
  • repaired/ — the same set of files scored against the coherence-repaired reference.

The Hugging Face dataset is data only; the scoring_kit/ used below is the one you cloned in step 1a. The source ontologies are not re-hosted — obtain each from its original publisher (see ontologies, and the dataset’s own ontologies.md).

2. Sanity-check your copy

Run the self-check against your downloaded data before you do anything else — it builds oracle submissions from the public splits and confirms they score perfectly:

python3 scoring_kit/self_check.py --data ./bio-ml

3. Subtrack 1 — Global equivalence alignment

For each pair, produce one alignment file using full OWL IRIs. The setting is semi-supervised: refs_equiv/train.tsv is public for tuning, but the test reference is hidden and scored organiser-side (there is no public global scorer). Validate the structure locally before submitting:

python3 scoring_kit/validate_global.py my-ncit-doid.rdf

Submissions are scored against the repaired, coherence-aware reference, plus reasoner-checked Global Coherence - see evaluation metrics.

4. Subtrack 2 — Local equivalence ranking

For each pair, read the gold-stripped candidate pools in bio-ml/<PAIR>/local.test.cands.tsv and emit a ranking of each query’s candidates, best-first. Validate the format against the pool, then self-score on the gold-bearing validation pool (local.valid.cands.tsv, whose TgtEntity column is the gold — usable directly):

python3 scoring_kit/validate_ranking.py bio-ml/NCIT-DOID/local.test.cands.tsv my-ranking.tsv
python3 scoring_kit/score_local.py my-ranking.tsv bio-ml/NCIT-DOID/local.valid.cands.tsv

Local ranking is scored with MRR and Hits@{1,5,10}, macro-averaged over the three pairs.

5. Submit

The evaluation window runs from 12 July to 1 September 2026 (00:00 Anywhere on Earth). Each scored subtrack has its own CodaBench competition:

Register on the relevant competition, then upload your submission as described on its Overview page. Results are published as provisional to the leaderboard; organisers verify, reproduce where possible, and mark accepted results alongside the organiser-run baselines.