We are currently using folder names like 'model_v2_final_final' and it is becoming a disaster. How do professional teams handle model versioning? Should the versioning be tied to the git hash of the code that created it? What about the data version? I need a robust strategy to make our experiments reproducible in our agile pipeline. What works best for you?
Effective model versioning requires using specialized tools like MLflow or DVC to track the specific combination of code, hyperparameters, and data versions used to generate each model artifact.
1 answer
Stop naming folders manually and implement MLflow or DVC immediately. Tying your model to a git hash is baseline hygiene, but you also need to track the data state and hyperparameters to actually achieve reproducibility. If you are still using 'final_final', you are just guessing, not doing data science.
Thanks for the advice, Dhanush Saniel. I’ve definitely been guilty of the 'final_final' thing myself, and I get a bit anxious thinking about how much I need to learn to fix it.
I definitely need to move away from my manual folder mess, Dhanush Saniel. This is a huge priority for me right now as things are getting way too chaotic to track.


I’m sorry if this is a silly question, Dhanush Saniel, but I feel a bit lost. I’ll look into MLflow, though I’m just worried about how much time it takes to set up.