RLHF-Reward-Modeling icon indicating copy to clipboard operation
RLHF-Reward-Modeling copied to clipboard
Published 8 months ago • verified publisher icon RLHFlow

Missing code for ODIN

Open maoliyuan opened this issue 1 year ago • 1 comments

In ODIN's code, args.correlation_with_length appears in the README.md, but it doesn't appear in other Python scripts or bash files.

Did the author forget to push the code of this part into this repo? I'm wondering about the detailed implementation of calculating Pearson correlation on the entire batch. This part of code seems to have relationship with args.correlation_with_length and is also missing in this repo.

Below are the calculation of Pearson correlation mentioned in ODIN's paper. Thanks a lot for answering! odin

maoliyuan avatar Dec 01 '24 07:12 maoliyuan

Sorry for the confusions! Please check the version here: https://github.com/Lichang-Chen/RLHF-Reward-Modeling/blob/odin/odin/gemma_two_head.py. I also pull request to merge the code into the main branch. Let me know if you have further questions.

Lichang-Chen avatar Dec 09 '24 20:12 Lichang-Chen