Great article! I follow a similar approach to keep rich output from each ML experiment I run. To keep my notebooks small I split my pipelines in several tasks so the training notebook only has the relevant parts. I created this tool to manage Data Science workflows and it supports running notebooks (uses papermill under the hood). I’d love to hear your thoughts: https://github.com/ploomber/ploomber

Here’s a code example: https://ploomber.readthedocs.io/en/stable/auto_examples/reporting.html#sphx-glr-auto-examples-reporting-py