search

mingkaid / rl-prompt

Accompanying repo for the RLPrompt paper
MIT License
286 stars 52 forks source link

The data in 'data/prompt-gpt2-vocab' folder seems meaningless? #4

Closed Dinxin closed 1 year ago

Dinxin commented 1 year ago

train.source.positive: there is only "LABEL_1".

WeChatWorkScreenshot_9bcd7291-95b1-43f1-9495-2c094d0b5925

train.target.negative: there are only repeating "issues"s. I don't think it is a meaningful input.

WeChatWorkScreenshot_ed0c285e-6a9e-45a6-829a-39d5feaa28b5

Could you please provide the instructions on the meaning of these files?

mingkaid commented 1 year ago

Thank you for your interest! These are placeholders for the prompt generator, and will be removed in future updates.

Specifically, the "LABEL_1" in train.source.positive is only used for text style transfer. "LABEL_1" refers to the target style label, which means positive sentiment for Yelp, and modern English for Shakespeare. In train.source.negative, you will see LABEL_0, which serves the same purpose.

The repeating "issues"'s in train.target.negative is likewise just a placeholder for generating prompts with 5 tokens.

I hope this answers your question. I'm closing this now because it's a clarification question. If you have any specific questions about your use case, please feel free to create another issue with the details, so we can assist