dnagpt/omnigene4-cpt-corpus

OmniGene-4 CPT corpus

Continued-pre-training (CPT) corpus for OmniGene-4 (see https://github.com/maris205/omnigene4 ). Total ~96 GB across DNA, protein, structure, and English-text replay splits.

Files

File	Size	Source / Description
dna_32g.txt	31 GB	DNA sequences sampled from public genomes
protein_uni_16.txt	16 GB	UniRef-derived protein sequences
protein_lucaone_15g.txt	15 GB	Protein sequences from the LucaOne pretraining pool
openwebtext.txt	37 GB	English text replay split (sampled from Skylion007/openwebtext); kept here so the entire CPT mixture is reproducible from a single repo
pdb_aa.fasta	100 MB	PDB amino-acid sequences (paired with the 3Di split below)
pdb_3di.fasta	100 MB	PDB Foldseek-3Di sequences (paired with pdb_aa.fasta)
ss.txt	263 MB	DSSP secondary-structure labels

Mixing ratio used in training

CPT runs on a 1 : 1 : 1 split of DNA : protein : OpenWebText, with the 3Di + DSSP files used for the dual-head per-residue classification objective in v5. See the OmniGene-4 paper (Methods §4.3) for the exact recipe.

Citation

@article{wang2026omnigene4,
  author    = {Wang, Liang},
  title     = {{OmniGene-4}: A Unified Bio-Language MoE Model with Router-Level
               Interpretability and Modality-Invariant Transfer},
  year      = {2026},
  journal   = {bioRxiv},
  doi       = {10.64898/2026.05.12.724542}
}