FishCaduceus/FishCaduceus-Pretraining-512

FishCaduceus Pretraining Dataset 512

Dataset description

This dataset contains fixed-length 512-bp genomic sequence windows prepared for pretraining FishCaduceus DNA language models. It uses single-nucleotide tokenization and is intended for masked language modeling.

Dataset summary

The dataset contains 6,087,221 JSON Lines records. Each record contains genomic coordinates and a seq field. The six files were audited by streaming line-by-line parsing; all records were valid JSON and all observed sequences were 512 bp long.

Source genomes

The sequences were derived from eight fish reference genomes, including seven cyprinid aquaculture fishes and zebrafish:

• Hypophthalmichthys nobilis
• Hypophthalmichthys molitrix
• Mylopharyngodon piceus
• Cyprinus carpio
• Megalobrama amblycephala
• Ctenopharyngodon idella
• Carassius gibelio
• Danio rerio

Dataset construction

The corpus consists of fixed-length 512-bp genomic sequence windows with species-aware and repeat-content-aware stratified sampling. Sequences are represented with single-nucleotide tokens for masked language modeling.

Dataset structure

Each JSONL record includes fields such as assembly, chrom, start, end, strand, and seq. The seq field is the nucleotide sequence used for pretraining.

Dataset splits

Split	Number of sequences	Sequence length
Train	5,782,509	512 bp
Validation	152,607	512 bp
Test	152,105	512 bp

Data characteristics

All final files in this repository were non-empty and readable as UTF-8 JSONL. The audit found no blank lines, invalid JSON records, missing seq fields, or non-ACGTN sequence characters. File sizes are recorded in FILE_MANIFEST.tsv at the upload-preparation root.

Intended uses

The dataset is intended for research on DNA language modeling, representation learning, and downstream analysis of fish genomes. Users should evaluate whether the sampling strategy and reference assemblies are appropriate for their application.

Limitations

The data represent selected reference genomes and sampled genomic windows; they do not constitute a complete representation of fish genetic diversity. Coordinate conventions, assembly versions, and sampling strata should be considered when interpreting results.

Related models

• FishCaduceus-20L-512
• FishCaduceus-28L-512

Citation

The FishCaduceus manuscript is in preparation. Citation information will be added after publication.

Acknowledgements

FishCaduceus was developed for research on fish genomes at the Institute of Hydrobiology, Chinese Academy of Sciences.

Contact

Xiao-Qin Xia
Institute of Hydrobiology, Chinese Academy of Sciences
Email: xqxia@ihb.ac.cn