search

TsinghuaC3I / SoRA

The source code of the EMNLP 2023 main conference paper: Sparse Low-rank Adaptation of Pre-trained Language Models.
62 stars 8 forks source link

Confirm the random seed numbers #3

Open SEONHOK opened 9 months ago

SEONHOK commented 9 months ago

Hi! The results in your paper are obtained from the random seeds "0 21 42 81 100"? As I saw the code, there are random seeds numbers "0 21 42 81 100". I just wonder whether it is correct or not.

I really appreciate any help you can provide!

telxt commented 9 months ago

Hi! You're right. We did use those five random seeds across all the experiments that required replication. Please don't hesitate to contact us if you have further questions or need more assistance!

SEONHOK commented 9 months ago

Thanks for the prompt reply! I have a few quick questions!

  1. I wonder whether you measure the performance of SoRA by the best performance over the checkpoints (in several check points M's in algorithm 1)
  2. Also, in algorithm 1, how do you measure "TrainUntilConvergence"? Is there any specific metric for this?

Many thanks in advance!