How to Preprocess Seed Data for Instruction Evolution to Generate conversation_0、Caption, Det fields

[yszhli]

Hi

I'm currently working on instruction evolution and would like to ask how to preprocess the seed data before running the evolution process. Specifically, I want to understand how to generate fields conversation_0, caption, det, etc., from the seed data.

Could you please provide some guidance or examples on the recommended preprocessing steps?

Thanks!

[RainBowLuoCS]

conversation_0 is the seed data of the evolved data
caption, is the meta data of the original data
det is the existed object position and class of the original data, you can use detetor like yolox to obtain this