Could not find the exact place where the training set is input?

[Dinxin]

This repo seems to only provide the prompt dataset in data folder, ignoring the needed training set (Yelp). I cannot reproduce the experimental results demonstrated in the paper. Could you please provide the full data and program which are necessary for the result reproduction?

[mingkaid]

Thanks for telling us about your issue, and I apologize for your experience.

The training data is in the prompt_tasks folder. For text style transfer, we use the preprocessed data which follows the tokenization format of GPT-2.

We will clean up the code further in the next update. I’m closing this now because it’s a clarification question

[Dinxin]

It seems that no programs read data from the prompt_tasks folder.

[mingkaid]

The file sql/rewards.py reads the data. You are also welcome to check out our restructured code (still WIP), which is located at the modular-interface branch

[Dinxin]

Many thanks. I check the restructured modular-interface branch.