---
title: MulinLab | hEvo
emoji: "🧬"
colorFrom: blue
colorTo: blue
sdk: static
short_description: Population-aware DNA modeling for human genomics
pinned: false
---

<div align="center">
  <img src="https://raw.githubusercontent.com/mulinlab/hEVO/main/header.png" alt="hEvo: learning human genetic variation across populations and evolution" width="100%">
  <h1>MulinLab</h1>
  <p><strong>Population-aware DNA modeling for human genomics</strong></p>
  <p>We study how human genetic diversity and haplotype sequence variation inform genome representation, variant effect, regulatory annotation, ancestry, and evolution.</p>
  <p>
    <a href="https://github.com/mulinlab/hEVO"><img src="https://img.shields.io/badge/GitHub-hEVO-181717?style=flat-square&logo=github" alt="hEVO source code on GitHub"></a>
    <a href="https://huggingface.co/datasets/MulinLab/hEvo-benchmarks"><img src="https://img.shields.io/badge/Hugging%20Face-Benchmarks-FFD21E?style=flat-square&logo=huggingface&logoColor=000" alt="hEvo benchmark dataset on Hugging Face"></a>
  </p>
</div>

---

## Explore hEvo

| Resource | What it contains | Access |
|---|---|---|
| [hEVO-base-o](https://huggingface.co/MulinLab/hEVO-base-o) | 87.1M-parameter, 1,024-nt model trained with masked language modeling on public 1000 Genomes Project and HGDP sequences. | Manual approval for peer reviewers |
| [hEVO-large-o](https://huggingface.co/MulinLab/hEVO-large-o) | 319.5M-parameter model in the same public-data training family. | Manual approval for peer reviewers |
| [hEvo-benchmarks](https://huggingface.co/datasets/MulinLab/hEvo-benchmarks) | 15 task bundles with frozen inputs or source downloaders, task-specific build scripts, and documentation. | Manual approval for peer reviewers |
| [GitHub source repository](https://github.com/mulinlab/hEVO) | Training and benchmark workflows, configurations, result tables, figures, and reproduction guides. | Public repository |

The `-o` suffix identifies a public-data training variant. It does **not** mean that model weights are openly licensed. Model-weight licensing has not been confirmed. Benchmark files may carry different terms; follow the cited source publication and task documentation for each dataset.

## How to get started

1. **Choose a model or benchmark task** from the table above.
2. **Request access** on the corresponding Hugging Face page. Model and dataset access is manually reviewed for peer-review use.
3. **Use the card's quick start** after approval. The model cards include DNA embedding and single-nucleotide likelihood examples; the dataset card lists task-specific download and build instructions.
4. **Reproduce the workflows** with the [GitHub guides](https://github.com/mulinlab/hEVO), which keep model code, evaluation workflows, and source result tables together.

## Scientific context and citation

The [GitHub repository](https://github.com/mulinlab/hEVO) describes the hEvo training stages, evaluation protocols, source tables, and figure workflows. For hEvo citation details, see the repository's [Citation section](https://github.com/mulinlab/hEVO#citation).

For questions about access or scientific use, contact MulinLab through the relevant repository's Hugging Face discussion page.