I am revising the model to solve QA task..

[0525stone]

Hi, I am working on your code to solve QA task.
I have a question.

Currently my dataset consists of context, question, answer(each question have pair of answer & context).
The length of question is too short(mostly, after tokenizing with 'bert-base-multilingual', 6 to 7 tokens are generated). so I add to make model run because default model setting requires 'chunk_size = 64'.

here is my question.
I think the role of token is just added to fill out empty space. I think that means logically is nothing.
Once training, I made two chunks and put those in. First chunk is made by "question: blahblahblah" and second chunk is made by "answer: blahblah"(those sentences are too short.. as I mentioned earlier)

Once I put input such as "question: blahblahblah? answer: " into the checkpoint model, the model spit out [CLS] question : blahblahblah? answer: [SEP] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD] [PAD].
I have no clue why my model stay dumb...

[yerinNam]

I'm having the same issue, did you solve it?