Zyphra/Zyda-2 · About annealing process

ldwang

Nov 1

Thanks a lot for your great dataset.
I have a few questions about annealing process used in evaluation.

Was the optimizer state loaded when the annealing phase started?
Since model A was used for annealing, does datasets comparison reflect that which dataset has a greater supplementary effect on model A?

Thanks for your reply.

Zyphra org Nov 5

Hi! Thanks for trying it out!

No, the optimizer was reset. This is part of the reason why the aggregate score initially drops during the annealing process (see Fig. 7 in our technical blog).
This is an interesting observation. From our practical purposes, we are interested in improving our own model A as much as possible, so running experiments using it makes the most sense for us. Annealing setup is very useful for that purpose, as it allows to see the signal on your own model on harder benchmarks (like MMLU) sooner, than running data experiments from scratch.

yury-zyphra changed discussion status to closed 14 days ago