Traing loss decreases but dev accuracy gives zero

Hi, I’m using T5ForConditionalGeneration model with Huggingface Trainer class to find relationships between words in a sentence such as street sign with black writing → (writing - black), (sign - street), (sign - with - writing)

I think training loss decreases when the model trains (which means it trains well…?), but it gives zero accuracy on eval, as shown in below figure…

So I wanted to check the outputs of training steps but don’t know how to…

How can I check the outputs of training steps…? and what might be the cause?

Thank you in advance!