Andrea Madotto

Oh, I didn't check that file, but you are right, no need, GPT tokenizer does the job already. Maybe open another issue? but good to know. Andrea

@jsedoc so in my decoding script I use multinomial sampling so the output is a bit different every time. If you want to try pure greedy use top-k 0 and...

Hey @nicolas-ivanov, yes I tried and yes it breaks the models' output. I believe that the model has not been trained using this positional token. Maybe because the model was...

Hi, Thanks for your repo. I have got the same problem here. Andrea

Maybe you need to provide a little bit more details. This could happen for many reasons: - the train_iter or val_iter is empty - the lr is too high or...

mmm I see. I really don't know at this point. I suggest going step by step inside the do_evaluation function to check where the loss gets none. Sorry I cannot...

Have you downloaded the BERT pre-trained model? https://drive.google.com/file/d/1Qawz1pMcV0aGLVYzOgpHPgG5vLSKPOJ1/view?usp=sharing

yes, I mean the pytorch_model.bin which yes should be inside the nli_model folder. So, we never had this error, could you add print(model) before model.to(self.device) this shouldn't be None. Maybe...

Hi, the KB to compute Entity F1 Score is used to generate the FP entities, together with the whole list extracted from the ontology. So, in general, you need 1)...

Hi, no currently works just in english. But I am aware of a GPT-2 trained on a Chinese corpus, so if you prepare some Chinese prompt I suspect it will...