metadata
license: mit
library_name: peft
Training procedure
Finally, it looks like overfitting has been circumvented.
Train metrics:
{'eval_loss': 0.11367090046405792,
'eval_accuracy': 0.961073623713503,
'eval_precision': 0.3506606081587021,
'eval_recall': 0.9097597679932995,
'eval_f1': 0.5062071663690367,
'eval_auc': 0.9359920115129883,
'eval_mcc': 0.5513080553639849}
Test metrics:
{'eval_loss': 0.11328430473804474,
'eval_accuracy': 0.9604888971537066,
'eval_precision': 0.34630886072474065,
'eval_recall': 0.9135862937475725,
'eval_f1': 0.5022370749476722,
'eval_auc': 0.9375606817360377,
'eval_mcc': 0.5489185177475369}
Framework versions
- PEFT 0.5.0