DS-UFT-Benchmark: Data supporting the evaluation of domain-specific unsupervised fine-tuning of ESM-2 650M

Update notice: A newer version of this archive is available. For the latest release and the version history, please visit https://doi.org/10.5281/zenodo.23045593. The updated archive supporting the revised manuscript, including the adopted trained prediction-head and SFT checkpoints, is available at https://doi.org/10.5281/zenodo.23050901. This earlier version remains available for provenance. This dataset contains the results supporting the revised DS-UFT-Benchmark manuscript, which evaluates full-parameter domain-specific unsupervised fine-tuning of ESM-2 650M. The archive includes sequence pseudo-perplexity (PPPL) results, contact-prediction evaluations, secondary-structure prediction results, GB1 mutational-effect prediction results, and the supplementary controls adopted in the manuscript. It provides result tables, selected per-protein or per-seed data, confidence intervals where reported, manuscript figure assets, and file manifests with SHA-256 checksums. GB1 results are restricted to validation-selected configurations. Regional PPPL results in Supplementary Table S4 use a consistent GA-filtered domain definition and include 19 proteins. Unused exploratory experiments and historical test-selected alternatives are excluded. The archive contains processed results and verification code, rather than a complete end-to-end training distribution. Model weights and complete third-party sequence and structure databases are not included. Please consult the README and manifests for analysis-specific scope, provenance, and limitations. Associated code repository:https://github.com/TianBoxue-lab/DS-UFT-Benchmark