search

mingkaid / rl-prompt

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

Scope of this project #24

Closed YujingYang666777 closed 1 year ago

YujingYang666777 commented 1 year ago

Hello,

I understand that this paper aims to create discrete prompts for LM and this is one motivation to optimize discrete prompts for black box models. So in this case, this method is applicable for GPT-3 as well, with some modification on the model part. Is that correct?

I have another question about the scope of the project. There are two experiments, 1. few-shot text classification and 2. Unsupervised Text Style Transfer. However, I am wondering if I can apply the project to the task of question answering. If so, the reword function should encourage prompts to answer the question correctly or closely. Is that applicable?

Thanks!

MM-IR commented 1 year ago

Hi, thank you for the questions! 1) That is true, with our current hyper-parameter engineering, in most applications with various models, as long as you can construct right reward function, usually with model logits, you can reach good performance for prompt optimization; 2) Yes, but adapting to other applications requires you to think about the correct formulation.

MM-IR commented 1 year ago

Since it is a clarification, I will close this issue. Thanks!