Closed tianyu-z closed 3 years ago
Small changes requested I think it would make more sense for you to walk us through the process to criticize it constructively :) Do you have reference experiments to look at how much v2 would help?
The experiment is here. The entropy_split term for this experiment is 0.67 which means 67% of the original sim data will be in easy_split.json and 1/3 will be in hard_split.json. I have run entropy_split>0.67 and it seems 0.67 is better than other cases (by my eyeballing). Entropy_split<0.67 is still running. I am able to give the result tomorrow.
Closing for now since we won't use it
Small changes requested I think it would make more sense for you to walk us through the process to criticize it constructively :) Do you have reference experiments to look at how much v2 would help?