ytyz1307zzh / auto-instruct Goto Github PK

Hello,

Thank you very much, and congratulations on your well-written code and paper.
I have a few questions regarding the training data of FlanT5:

1. As I understand it, you classified the tasks using the 'category' field in each 'task_metadata.json' file, found at 'data/niv2_english/TASK_NAME/task_metadata.json', resulting in 575 tasks for training. Is this correct?
2. For each task, you sampled 400 examples (pairs of candidate instructions, x, y, plus RougeL), following the filters outlined in Appendix A. You mentioned ending up with a total of 122K examples for training, but with 400 examples for each of the 575 tasks, this would amount to 230K total examples. Are there tasks with significantly fewer than 400 examples?

Hello,how can i use my own openai-key in gpt-utils.py, i always get HTTPError:500 in result.
When I replace the openai.api in the "Authorization: Bearer" line of the script with my own key, the message in response is still "You didn't provide an API key. You need to provide your API key in an Authorization header".
how can i deal with it?
Thank you very much!!!

I still wonder why it needs to train a specific model to rank the instructions? Can GPT-4 be the ranker with a prompt templete?

Hello, I am very interested in your work. Regarding the instruction generation part in section 4.1, I would like to ask if the instructions generated using the black-box LLM can be made public? I would be very grateful for your reply to my question.

Hi @ytyz1307zzh ,
Is it possible for you to release the trained ranker model as soon as possible?