Data and downloads
Download the scores shown on the leaderboard, including the original v1 rows and the controls added in v1.1. The catalog below lists the task inputs and analysis files behind those scores.
Leaderboard downloads
Benchmark file catalog
The research repository is private. The gene-level task inputs, scripts and provenance tables listed below are not hosted on this site. A Zenodo deposit is in preparation; it has no public DOI yet. For questions about access, contact the corresponding author.
Provenance tables
provenance_human_protein_coding.tsv has one row per gene: symbol, entrez_id, exposure (PubMed articles linked in gene2pubmed, counting articles with at most 20 genes), exposure_bin, genept_key, genept_2023_source and genept_2023_tag_date (the tag in GenePT's October-2023 text), genept_2024_template_flag (detector on GenePT's untagged 2024 text) and ncbi_2026_source (tag of the NCBI Gene summary on 6 October 2026). Source values: alliance (generated from GO), refseq (curated), other, none, absent. provenance_genept_keys.tsv carries the same flags for all 33,703 GenePT keys.
Template detector
Python standard library only. It reads the [provided by ...] tag when present and otherwise flags a summary whose first sentence starts with an Alliance template phrase; on GenePT's 2024 text against the 2023 tags: precision 1.000 and recall 0.999 for the revised detector. The original template list had recall 0.997.
from gene_text_audit import is_generated, strip_generated
summary_filtered = {gene: strip_generated(text) for gene, text in summaries.items()}
Licences
Derived labels and tables: GO annotations (GO Consortium, CC BY 4.0); DepMap 24Q4 (CC BY 4.0); gnomAD v4.1 (CC0); ClinGen (CC0); Human Protein Atlas (CC BY-SA 3.0); GTEx (open access). Only derived binary labels and term matrices are redistributed, with attribution. The prepared deposit declares CC BY 4.0; upstream source terms still apply, including the HPA attribution and share-alike terms. A repository-wide code licence is not supplied in this checkout. GenePT text (Zenodo 10833191) and NCBI Gene summaries are not redistributed; the corpus builder applies the provenance flags to your own copy.
Metadata correction of 25 October 2026
The exposure column of both released gene tables now contains the filtered publication count that defines exposure_bin; the original export had written the unfiltered count into it (19,141 values changed; bins, folds, labels and every stored score were unchanged). Before-and-after checksums and a regression check are in review_2026-10-25/ of the repository.