Pre-train models do not perform as well as they should

[Dinple]

Hi,

I am following the pre-train instructions provided here.

I have 7 almost identical netlists (the same #macros) for the pre-training set and 1 target netlist (the same #macro) for the evaluation.
I launch the training job with the 7 netlist.
I launch the replay buffer with --num_netlists=7.
For the collection jobs, each netlist is given an equal amount of resources (30+ collectors).
I launch one eval job on the target netlist.

However, the pre-train model does not converge after 200 iterations. I also tried to fine-tune the pre-train model, but the fine-tuned model also does not converge after 200 iterations.

Could you identify what I am doing wrong, or are there some settings I need to tune for pre-training?

Thanks!