search

hanjanghoon / BERT_FP

Fine-grained Post-training for Improving Retrieval-based Dialogue Systems - NAACL 2021
97 stars 19 forks source link

when Post-trained douban dataset the loss function shake little,is that correct? #4

Open andrew-begain opened 2 years ago

andrew-begain commented 2 years ago

dear hanjanghoon thanks for you greate work! when i Post-trained douban dataset the loss function shake around 2.0 ,is that correct?

hanjanghoon commented 2 years ago

I remember that the loss during post-training was about 1.7~2.0, and about 0.2 during finetuning.

andrew-begain commented 2 years ago

I remember that the loss during post-training was about 1.7~2.0, and about 0.2 during finetuning.

@hanjanghoon thanks for your reply, How does this algorithm apply to a real conversation system? It's a question and a answer format .thank you very much