Clarification in the paper

[LeeShiyang]

Hi authors, thank you for releasing code and data for this project. I am confused about the following part in the paper.

For fair comparison, we replace Alpaca’s original Davici-003 response with ChatGPT’s response, and also sample 70K instructions subset to train WizardLM.

(1) In your alpaca baseline, do you use their original data or response generated from ChatGPT? (2) Since you use the dataset from alpaca as seed dataset, do you change original alpaca response to train WizardLM? (3) What does the following sentence mean? My understanding is that you sampled 70k data from the full 250k data, which have already included alpaca' seed dataset. Why do you say also sample 70K instructions subset to train WizardLM.?

also sample 70K instructions subset to train WizardLM.

Thank you for your clarification.

[murali1999-tech]

1. In our alpaca baseline, we used Alpaca's original data and their original Davinci-003 response. We did not replace their response with one generated by ChatGPT for this baseline.
2. No, we did not change the original Alpaca response to train WizardLM. In our experiments, we used the same Alpaca data as our seed dataset and did not modify the responses.
3. You are correct that the full Alpaca dataset already includes their seed dataset. However, we sampled a subset of 70K instructions from the full Alpaca dataset to train WizardLM. This was because the full dataset was too large for our computational resources, and we found that a subset of this size was sufficient for training. So when we say "also sample 70K instructions subset to train WizardLM," we mean that in addition to using the Alpaca dataset as our seed dataset, we also sampled a subset of 70K instructions from the full dataset to train WizardLM.

[haiatn]

If all is well can this issue be closed?