Validation Loss as well as validation F1 increasing
Abierto
- Lenguaje dominante
- Python
- Estrellas
- 1.7k
- Forks
- 232
- Métricas de merge de PR
- Sin PR fusionados en 30 d
Descripción
Hi,
While finetuning scibert on JNLPBA and bc5cdr dataset, I observed very strange phenomenon.
The validation loss is increasing whereas the validation F1 in either increasing or staying the same.
Please let me know because I have faced a similar issue when I used pytorch-crf over scibert. There also for ner (JNLPBA and bc5cdr) validation loss anf validation f1 are increasing.
Guía de contribución
No hay ninguna guía de contribución indexada para este repositorio
Evaluación
Este issue todavía no se ha evaluado.