KeyError: 'logits'

Hi,
I have created and trained a tokenizer. I then loaded using

tokenizer = PreTrainedTokenizerFast.from_pretrained(‘bert-base-he’)

I then created a model using starting to load
model = BertForPreTraining.from_pretrained(‘bert-base-uncased’)

I did everything to train the model both with the MLM approach and the NSP approach and everything worked fine. So I saved the model in my google drive.

I now want to use the model but I get the following error: KeyError: ‘logits’

This is the code I use:

Initialize MLM pipeline

from transformers import pipeline

mlm = pipeline(
“fill-mask”,
model = torch.load(‘BHSA01.pt’, map_location=torch.device(‘cpu’)), #it seems it wants just the model name in .json format
tokenizer = PreTrainedTokenizerFast.from_pretrained(‘bert-base-he’)
)

Get mask token

mask = mlm.tokenizer.mask_token

Get results for a particular masked phrase

phrase = f’{mask} בְּ רֵאשִׁ֖ית בָּרָ֣א אֱלֹהִ֑ים אֵ֥ת הַ שָּׁמַ֖יִם וְ אֵ֥ת הָ’
result = mlm(phrase)

Print result

print(result)


Just one last note. Biblical Hebrew is written right to left. Can that create problem?

Thanks a lot for Your help!

Elia