kristiyanvachev / leaf-question-generation Goto Github PK

Hi

Thank you very much for your implementation. It's been very helpful to me.
Regarding your example script to generate questions and answers (google collab), I would like you to clarify some doubts for me, if possible.

1 - Do you know if it is possible to get generated tokens in training_step? Instead of getting them just at the end, via generate() inference method.

2 - Do you have a particular reason to do encoding using tokenizer(answer + <sep> + context, ...) instead of using tokenizer(answer, context, ...)?

3 - Have you encountered overfitting throughout your experiments? Unfortunately, my dev loss only improves up to the second epoch (cross-entropy loss of 1.35) and then increases. From what I observer from your experiments, you can reach the 4th epoch with a loss of 1.17374.
Note: I am using the same SQUAD v1.1 splits, model: t5-base, batch size: 32, optimizer: AdamW (eps = 1e-6).

4 - Related to the previous question. Do you have any idea what a "good loss" is for this QG task? In any case, I haven't seen the dev loss reach a value lower than 1.

Thanks in advance.
Bernardo

Hi Kristian,

Thanks for the code, it is really helpful.
I'm wondering about one thing: how do you ensure that your distractor module will output a sequence with 2 tokens in it? Because it appears to me that training with three distractors doesn't necessarily make you 100% certain that the model will generate outputs of this form. Or do you do multiple samples and just concatenate them?
Thanks in advance.

Hi All,
I am working on multi response answer generation model using nlp.
Please anyone help me

Hello
Could you please tell what python version did you use ?

Hello,
First of all, amazing work! I have so much to learn!

I apologize if the question is too basic: I followed all the steps of installation but I'm not sure how I can actually "run" the code in a way that the Angular app can send requests to it.

Thank you very much!