PhenoNormBench

PhenoNormBench is a mention-level benchmark for normalizing free-text biomedical phenotypes to Human Phenotype Ontology (HPO) identifiers. This release combines seven standardized datasets into one file with 13,390 records. It contains only a test split; no training or validation split is provided.

Load the Dataset

from datasets import load_dataset

dataset = load_dataset("songjie0209/PhenoNormBench")
test = dataset["test"]
fgdd = test.filter(lambda example: example["dataset"] == "FGDD")

The dataset column identifies the source of each record. All seven sources are included in test by default.

Dataset Composition

Source (dataset)	Records
GeneReviews-10	352
ID-68	866
GSC2017	2,773
GSC2024	2,910
CSC	1,783
FGDD	1,866
BC8_T3	2,840
Total	13,390

Each record has the same six fields:

Field	Type	Description
dataset	string	Source dataset name.
source_row	integer	One-based row number in the source Excel sheet, including the header row; the first data row is 2.
Raw_Phenotype_Names	string	Original free-text phenotype mention.
Standard_Phenotype_Names	string	Standardized phenotype name.
HPO_id	string	Reference HPO identifier.
Accepted_HPO_ids	list of strings	HPO identifiers accepted for evaluation, including HPO_id.

Accepted_HPO_ids is a list in this release. In the source Excel workbooks it is stored as JSON text. Source-specific Patient_ID, patient_ids, and pmids columns are omitted from this unified release.

Source and Processing

The source workbooks are available in the OntologyAligner Dataset directory. Each workbook was read in its original row order.

Citation

The OntologyAligner manuscript is under review. The preprint is available at https://arxiv.org/abs/2609.10055.

License

The OntologyAligner repository is released under the MIT License. Check the original datasets for any additional source-specific terms.