Training steps needed to reproduce the reported number

Thank you for releasing DeepEyes! It is an excellent work and means a lot to the community!

I’m now attempting to reproduce the results reported in the paper, and although here suggests 80 steps in paper, I found using your script (7b version) cannot reproduce the results in the paper. (-3% in Vstar and -10%+ in HR bench) Could you clarify the exact steps used to obtain the published numbers? Thanks so much!