OmegaGenome: source code

Abstract

Source code accompanying OmegaGenome: Toward Better Sub-Million Scale Expert Models from Large Genomic Language Models via Knowledge Distillation (Science Advances): teacher fine-tuning, two-stage knowledge distillation, evaluation, the base-resolution multi-track regression pipeline, and the code that produces the paper's figures. Version 2.1 corrects the Zenodo DOI referenced in README.md and CITATION.cff to the concept DOI, which always resolves to the latest archived version. Version 2.0 superseded version 1.0, which had split the same material across three archives built from three branches. Everything is now a single tree that matches the public repository: the six main-text figures as submitted, together with the scripts that draw them and the result tables those scripts read (paper_figures/); 22 of the 23 figure scripts regenerate from the bundled tables with no GPU and no data download; the completed base-resolution multi-track regression pipeline (NTv3 teacher fine-tuning, track students, teacher-output caching and the size sweep), with unit tests; machine-specific paths replaced by documented environment variables, and the dependencies the code imports declared in pyproject.toml with a regenerated uv.lock. See README.md in the deposit for the contents and README.md inside the archive for installation and usage. No data were generated by this study: the 18 classification tasks are the revised Nucleotide Transformer benchmark on Hugging Face, and the regression benchmark is built from public ENCODE and GENCODE data. Training logs, intermediate feature arrays and model checkpoints are not included; the trained teachers and students can be regenerated from the public datasets with this code.