google-bert

/

bert-base-uncased

like 1.53k

Could it be that the weights of BERT base, or rather this model, are not suitable for the Yelp dataset (with lengths ranging from 180 to 400, and 5 classifications)? Currently, in my fine-tuning for sentiment analysis with 5 classifications, the accuracy can reach a maximum of only 60%. Has anyone else encountered such a situation?