hEvo pretraining and data-construction code
This repository contains the source code for the hEvo pretraining pipeline and its cohort-aware preprocessing statistics workflow. Access to files is manually reviewed by MulinLab. Sign in to Hugging Face and select Request access on this page; approved users can then clone or download the source.
Contents
hEvo_pretrain/: Stage I haplotype corpus preparation and masked-language-model training, plus Stage II variant-centred dataset generation and prediction-head calibration.hEvo_benchmark/scripts/preprocess_stats/: cohort genotype normalization, phased haplotype-window statistics, global sequence deduplication, validation, and Fig. 1 source-table generation.
This repository contains code and placeholder configurations. It does not contain source VCF/BCF files, reference genomes, sample lists, generated training datasets, model checkpoints, run outputs, or access credentials. Keep source data and generated artifacts on approved local or HPC storage and apply each source dataset's terms.
Access and setup
Sign in to your Hugging Face account and request access to this repository. MulinLab reviews requests manually.
After approval, install Git LFS if required by your environment, authenticate with
hf auth login, then clone:git clone https://huggingface.co/MulinLab/hEvo-pretraining-codeFollow the README in the selected workflow. The scripts target Linux/HPC environments; Slurm, C++/HTSlib, PyTorch, and other dependencies vary by stage. The example commands use the
llmbmmicromamba environment when available.
Public resources
- hEvo GitHub repository: public benchmark code, figures, and numerical source tables.
- Published preprocessing statistics: retained CSV results and SHA256 checksums.
- MulinLab models and benchmark data.
The source code repository does not currently declare a software license. Contact MulinLab before redistributing or reusing the code beyond the access granted here. Dataset and reference-genome terms remain specific to their original sources.