esm2_t6_8M_binding_sites_finetune_600K

ESM-2 Full Finetune for Binding Sites

This model is a full finetune of ESM-2, to illustrate how full finetuning overfits and generalizes quite poorly compared to LoRA and QLoRA finetuning. This model was finetuned on the 600K dataset. We also note that on the 24GB A10 GPU, the batch size has to be significantly smaller than when using LoRA or QLoRA. To finetune a similar model, use this script.

Overfitting

Train metrics:

{'eval_loss': 0.13651661574840546,
'eval_accuracy': 0.9656322509450104,
'eval_precision': 0.38616650354104665,
'eval_recall': 0.9618091516702236,
'eval_f1': 0.55107594226701,
'eval_auc': 0.9637635647574605,
'eval_mcc': 0.5977943918337999}

Test metrics:

{'eval_loss': 0.2910114824771881,
'eval_accuracy': 0.923270649115702,
'eval_precision': 0.14887069127765168,
'eval_recall': 0.533511928419524,
'eval_f1': 0.23278520670392827,
'eval_auc': 0.7327381144575454,
'eval_mcc': 0.25329082069818704}