How to use my own database for training and evaluating Retro for Question-Answering?

[Zahin112]

What is the main purpose or outcome of this training? is it only for text completion? How do I finetune it for Question-Answering or any other downstream task?

I want to use my dataset in the training. How do I do it and which format the dataset needs to be?

Is there script to evaluate the model reproducing the results of the paper?

[vpj]

Which model are you referring to?

[Zahin112]

Sorry I forgot to mention. I was referring to Retro.