transformer_pwff_lr3e-4

Positional-wise feed-forward (PWFF) pooled transformer encoder trained on concat [r, |r-a|, r*a] with r = ESM2 8B rbd_embeddings and a = ESM2 8B ace2_embeddings (mean pooled).

• Dataset: BIIE-AI/ace2_binding
• Training: 8 epochs
• Learning rate: 3e-4
• Reported average per-specie MCC: 0.762

Files

• transformer_best.pt: PyTorch checkpoint (best model)

Notes

This repository contains a raw PyTorch checkpoint only. To use it, load it in the same codebase/model definition that produced the checkpoint.