Converged to something similar after spending 2 days bissecting a repo to reproduce a training run, having to wait 3hr on each commit before conclusive results. I couldn't get myself to use hydra though, it felt like a lot of bloat vs loading a yaml with pydantic