Generation details for models under data/labeled about factscore HOT 1 OPEN

Hi @caiqizh, thank you for your interest in our work.

Here is the prompt we used for ChatGPT:

Here are two hyperparameters:

• temp=0.7 for both ChatGPT and InstructGPT
• max_tokens=512 for InstructGPT and max_tokens=1024 for ChatGPT

Using different max_tokens should not affect the generations unless the generation exceed max_tokens , which never happened in our case. Given this, I think it is possible that you are seeing much longer responses due to the internal changes in ChatGPT (if it's not due to the difference in the prompt).

Let me know if you have any further questions. Thanks.

from factscore.