README.md

metadata

language:
  - en
pretty_name: hEvo benchmarks
tags:
  - biology
  - genomics
  - DNA
  - benchmark
gated: true
extra_gated_heading: Reviewer access request
extra_gated_description: >-
  This repository is reserved for peer reviewers. Please sign in to Hugging Face
  and submit an access request with your review purpose. MulinLab will review
  requests manually.
extra_gated_button_content: Request reviewer access

hEvo benchmarks

15 reproducible task bundles for human population genomics

Frozen evaluation inputs or source downloaders, task-specific construction scripts, and documentation for peer review. Model weights are distributed separately.

Task catalog · Request access · Build data · GitHub workflows · hEVO-base-o · hEVO-large-o

---

Contents

• Dataset overview
• Task catalog
• Access and download
• Reproduce a task bundle
• Sources, terms and citation

Dataset overview

At a glance	Details
Collection	15 task bundles maintained by MulinLab for peer review
Contents	Frozen inputs or upstream downloaders, task-level build.py, and construction documentation
Model dependency	Model-independent dataset builders; model weights are hosted separately
Build environment	Python 3.10–3.12; validated with Python 3.12
Access	Gated; requests are reviewed manually
Data terms	Source-specific; this collection does not claim one license for all included data

Task catalog

Choose the task folder for its source descriptions, labels, splits, construction parameters, seeds and build command.

Task	Evaluation content	Task guide
adaptation	Selected-allele preference	Open guide
ancestry	Genetic ancestry	Open guide
causal_variants	Fine-mapped GWAS candidate variants	Open guide
clinvar	ClinVar snapshot dated 2026-06-15	Open guide
clinvar_reclassification	Retrospective ClinVar reclassification	Open guide
deepsea	DeepSEA	Open guide
denovo_variants	DDD de novo variant controls	Open guide
evolution	HAR and HAQER classification	Open guide
haqer_windows	HAQER grouped windows	Open guide
genomic_elements	Genomic elements	Open guide
mpra	MPRA emVar	Open guide
rare_causal_variants	Rare fine-mapped GWAS candidates	Open guide
rare_variants	UK Biobank rare variant associations	Open guide
ure_binary	URE binary classification	Open guide
ure_multiclass	URE multiclass classification	Open guide

Access and download

Access is gated for peer reviewers and approved manually by MulinLab. Sign in to Hugging Face, request access on this page, and wait for approval. After approval, authenticate locally with hf auth login using your own read token. Download only the task files needed for your analysis.

For example, to fetch the MPRA bundle and shared requirements file:

from huggingface_hub import snapshot_download

snapshot_download(
    "MulinLab/hEvo-benchmarks",
    repo_type="dataset",
    allow_patterns=["mpra/*", "requirements.txt"],
    local_dir="hEvo-benchmarks",
)

Individual data tables and source workbooks are also available from the repository Files tab. For one ancestry population or one ClinVar reclassification seed, download the corresponding data/*.tar.zst archive and unpack.py, then run:

python unpack.py ancestry --group afr
python unpack.py clinvar_reclassification --group seed_01

Files are extracted under the selected task's data/ directory. Omit --group to extract all downloaded archives for that task.

Reproduce a task bundle

Use Python 3.10–3.12 (validated with 3.12) and install the construction dependencies:

pip install -r requirements.txt
python <task>/build.py --help

Run the command documented in that task's README. Common path options are resolved against the corresponding script directory. Reference-based tasks require an external indexed hg38 FASTA and its .fai index. Add external tools only where needed: samtools to index a FASTA, curl for ENCODE, and bedtools/pybedtools for recreating the optional evolution negative pool. URE builds use the Python standard library. DeepSEA tensors and HAQER LMDBs are built locally.

Sources, terms and citation

Each task guide documents its source data, publications, labels, splits, output files, construction parameters and random seeds. The terms of the source publications remain applicable; approval to access this collection does not replace them. Check the task guide and cited source before using or redistributing any material. No single collection-wide license is claimed.

For code and complete benchmark workflows, see mulinlab/hEVO. For the evaluated checkpoints, see hEVO-base-o and hEVO-large-o.