Thank you! I'm quite happy with the core_eval.py rewrite. I wanted to evaluate my base model with the DCLM "core score" as described in their paper, but what felt like it should surely be a simple thing of ~300 lines of code actually required me to pip install and depend on a