Re-Align/URIAL

Scripts for experiments on "ALIGNMENT AS TOKEN DISTRIBUTION SHIFT"

Closed this issue · 5 comments

Sreyan88 commented

Hello!

Great work. Do you plan on releasing the code for the analysis you do in your paper in the subsection "ALIGNMENT AS TOKEN DISTRIBUTION SHIFT"? Thank You!

yuchenlin commented

Hey Sreyan,

Thanks for the question. Yes, I will clean the code and upload it here soon. Will ping you here once it is ready! :D

Best,
Yuchen

Sreyan88 commented

Thank You! Looking forward!

Sreyan88 commented

Hi @yuchenlin,

I am trying to replicate some of your analysis with different combinations of LLMs and training datasets, while you are getting your code ready. However, I am not sure how in the huggingface model.generate() function one can return a list of top-k tokens at each step of decoding. Can you please help me with this?

Thank You!

yuchenlin commented

Hey @Sreyan88,

I uploaded a (very rough) version of the code about token distribution analysis here: https://github.com/Re-Align/AlignTDS

Would u plz take a look? (I haven't got time to make better documentation and testing yet (will do soon), so please be prepared that there might be a few minor bugs).

Best,
Yuchen

Sreyan88 commented

Hi @yuchenlin ,

Thank You so much for the scripts! Had a look and looks promising! Again, great work!

  • 👍1