Benchmark artifacts and FlatfishBert model checkpoints for "A Governance-First Benchmark for Genomic Fragment Classification under Homology-Aware Evaluation"

This record contains the processed benchmark artifacts supporting the study “A Governance-First Benchmark for Genomic Fragment Classification under Homology-Aware Evaluation”. Contents include processed benchmark datasets for the governed core benchmark (Exp1–Exp5), processed diagnostic-task datasets (Task G, Task S, and Task N, including reverse-complement augmentation variants where applicable), processed cross-species evaluation sets, supplementary-table source files and mappings, similarity-audit artifacts, result summaries, and manifest/checksum files for verification and reuse. This record is intended to support manuscript review, result verification, and downstream reuse of the processed benchmark artifacts. It should be used together with the associated GitHub code repository, which contains the scripts, environments, and execution logic for benchmark generation, training, evaluation, figure generation, and MMseqs2-based similarity auditing. Raw NCBI RefSeq genome FASTA/GFF files are not redistributed in this archive. The reference assemblies and annotations are publicly available from NCBI RefSeq and are identified by accession in the manifest files. DNABERT-2 is used in the study as a strong public baseline. Fine-tuned DNABERT-2 checkpoints are not redistributed in this record. Instead, this archive provides the exact processed datasets, result summaries, and supporting artifacts required to verify the reported DNABERT-2 comparisons. If FlatfishBert checkpoints are included in this record, they are provided to support result verification and downstream reuse only. The associated code repository remains the primary source for runnable training and evaluation logic. Related resources:- Code repository: https://github.com/dogesang/FlatfishBERT - Code archive DOI: https://doi.org/10.5281/zenodo.19840124 - Data archive DOI: https://doi.org/10.5281/zenodo.19825273