dnagpt/omnigene4-mm-corpus

OmniGene-4-MM unified corpus

Multi-modal training corpus used for the OmniGene-4-MM Stages 1–3 (see https://github.com/maris205/omnigene4 ).

Each row is a JSON object with messages (chat-format), images (list of relative image paths), and modality field. Vision rows reference images that live in the source datasets:

• Vis-CheBI20 (PharMolix/Vis-CheBI20)
• PubMedVision (FreedomIntelligence/PubMedVision)
• HPA10M (Human Protein Atlas microscopy)
• ChartQA (HuggingFaceM4/ChartQA)
• Synthetic biomedical visual tasks (project-internal)

To reproduce the MM training, download the source vision datasets to matching local paths and run omnigene5/scripts/02-fix_image_paths.py to remap.

Files

File	Rows	Notes
train.jsonl	~280K	Used by 30-train_omnigene5_stage1.py (vision warmup) and the mixed-stage pools downstream
val.jsonl	~5K	Held-out validation slice
meta.json	—	Per-modality counts and split sizes
B_chebi20/train.json	~26K	Vis-CheBI20 train split (project copy of PharMolix/Vis-CheBI20). Five OCSU subtasks: struct_recog, struct_cap, general_desp, trans_iupac, trans_smiles.
B_chebi20/test.json	~3K	Vis-CheBI20 test split, same five subtasks. Used by 60-eval_stage2.py, 91-eval_stage3v2.py, 92-eval_stage3v3.py, 95-qualitative_demo_stage3v3.py.

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}
}