Using T5 for multi-label classification --> Possible to calculate LRAP?

[qsc-hamburg]

Hi,

I was experimenting with T5 and a four category multi-label classification were I would like to calculate also ranking loss metrics. Is it in any sense possible to extract the model outputs/softmax values after the evaluation? I would like to use them to calculate the LRAP.

Using the traditional transformers it was easy to extract them:
result, model_outputs, wrong_predictions = model.eval_model(eval_df)

Now with T5, I can't get the model outputs directly.
result = model.eval_model(eval_df)

Is it in any sense possible to access them?

Thank you for your help in advance!