ESMfluc: trained models for Protein Block variability prediction and signed contribution analysis

Trained ESMfluc model weights for residue-level single-PB versus multi-PB classification and exact signed-contribution analysis. Includes seven model conditions with three independent training seeds each: ESM2 and ESM3 frozen, top-four and top-28 fine-tuned BiLSTM--attention models, plus the frozen ESM2 linear baseline. Shared backbone tensors and seed-specific overrides reconstruct all 21 original checkpoints without quantization or precision conversion. Includes training arguments, class definitions, band-detection settings, checksums, a checkpoint loader and environment/source metadata used by the ESMfluc Colab notebook. Contributions describe the model's final calculation and do not establish causal biological interactions. Built with ESM.