ViroBland: A Small-Scale Mixed Pre-training Corpus

ViroBland (also referred to as ViroBlend) is a small (216 Mbp) mixed pre-training corpus that combines broad genomic context with enriched viral signals. It is built with source-wise stratified sampling to balance three heterogeneous data sources, enabling compute-efficient and reproducible pre-training while retaining strong viral in-domain coverage.

This dataset was introduced in the paper ViroBench: Benchmarking Nucleotide Foundation Models on Viral Genomics Tasks.

• Project Page: https://sii-agi4s.github.io/ViroBench/
• GitHub Repository: https://github.com/QIANJINYDX/ViroBench

---

Overview

• Scale: Small (216 Mbp; ~216M bases)
• Sources: 3 (Human reference + Multi-species + Viral in-domain)
• Final split sizes: Train 32,023 / Val 4,000 / Test 2,662 sequences

---

Data Sources

1. Human Reference (GRCh38 / hg38; following HyenaDNA)
2. Multi-species Diversity (Nucleotide Transformer dataset)
3. Viral In-domain Data (OpenVirus)

Citation

@article{ye2026virobench,
  title={ViroBench: Benchmarking Nucleotide Foundation Models on Viral Genomics Tasks},
  author={Ye, Dongxin and Hu, Fang and Hu, Han and Hu, Shu and Tan, Yang and Ouyang, Wanli and Li, Stan Z and Cui, Jie and Dong, Nanqing},
  journal={arXiv preprint arXiv:2605.25388},
  year={2026}
}