ViroBlend: A Small-Scale Mixed Pre-training Corpus for Genomics

ViroBlend is a 216 megabase pair mixed pre-training corpus introduced by YDXX. It combines broad genomic context with enriched viral signals using source-wise stratified sampling to balance three heterogeneous data sources. The dataset was last updated on 2026-05-29.

Use Cases

• Pre-training language models on genomic sequences based on the mixed corpus design
• Benchmarking model performance on viral in-domain tasks based on the enriched viral signals
• Studying the effect of stratified sampling on pre-training efficiency based on the source-wise balancing method

Strengths

• Corpus size is explicitly defined as 216 Mbp
• Designed for compute-efficient pre-training via stratified sampling of three data sources
• Aims to retain strong viral in-domain coverage

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download
• Column-level documentation is absent; field semantics must be inferred after download

Provenance

Source

YDXX via Hugging Face

Collection Method

Built with source-wise stratified sampling from three heterogeneous data sources

Freshness

Last updated 2026-05-29 08:33:58