Tips on improving QAG for a custom product related data set. [Not an Issue]

[amithadiraju1694]

Hey @asahi417 , great work , extremely useful set of models !

I have a custom data set which involves product related information, want to use your models for Q&A pair generation. So far, I found E2E QAG is performing better than Multi Task and Pipeline in generating meaningful Q&A pairs.

I tried different text prompt prefix strings , to my context which seemed to improve Q&A pairs a little, but , not by much. Do you have any additional tips on generating more meaningful and contextually valid Q&A pairs with these models, other than manipulating generate function like more beams ?

P.S: Example of what I tried prefix -> "This is a product metadata, generate Q&A pairs which regular shoppers would have " , context " Contains milk, eggs, shellfish fragments etc". Input if prefix + context to generate method.