Hi thank you for the project!
Im trying to reproduce it on the 3B baseline.
how much steps did you used for 3b coldstart SFT?
too much SFT decreases the performance. so im wondering how much steps should I try to make the performance on the paper.
Thank you!
Hi thank you for the project!
Im trying to reproduce it on the 3B baseline.
how much steps did you used for 3b coldstart SFT?
too much SFT decreases the performance. so im wondering how much steps should I try to make the performance on the paper.
Thank you!