metadata
license: cc-by-4.0
language:
- en
- zh
size_categories:
- 100K<n<1M
task_categories:
- text-generation
- question-answering
tags:
- biology
- protein
- DNA
- bioinformatics
OmniGene-4 SFT corpus
Supervised fine-tuning data for the OmniGene-4 / OmniGene-4-MM family. See https://github.com/maris205/omnigene4 for the training scripts that consume these files.
Files
| File | Rows | Used by | Description |
|---|---|---|---|
bio_sft_v2_train.jsonl |
~179K | Bio-SFT v2 | Eight task families: protein homology (BioPAWS), DNA, structure (3Di/DSSP), cell biology, molecules, mutation, structure prediction, general bio QA |
distill_seed.jsonl |
~6K | seed-only | Initial distillation seed used to bootstrap the v2 corpus |
train/omnigene_sft_v1_train.jsonl |
~179K | SFT v3 base | Same as v2 with cleaner schema and Alpaca template |
train/omnigene_sft_v1_train_with_remote.jsonl |
~199K | SFT v3-v5 (final) | Above + 20K BioPAWS protein_pair_remote rows; this is the file used by Bio-SFT v3, v4, v5, and the OmniGene-4-MM Stage 2/3 LoRA training |
eval/omnigene_sft_v1_eval.jsonl |
~1.5K | held-out eval | Used by 40-eval_omnigene4mm.py, 60-eval_stage2.py, 92-eval_stage3v3.py, etc. |
master/omnigene_sft_v1_master.jsonl |
~285K | (intermediate) | Pre-split master corpus before train/eval split |
master/cell_sft_master.jsonl |
~37K | task subset | Cell-biology SFT subset |
master/mol_sft_master.jsonl |
~50K | task subset | Molecule SFT subset |
stats/data_mix_report.json |
— | metadata | Per-category counts and ratios |
Schema
Each line is a JSON object with at least instruction, input, output,
and category. Some rows additionally carry task_name and subtask for
fine-grained accounting.
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}
}