Keep trained checkpoints out of git: pin the hash in a freeze file, store weights in object storage
Teammates often commit best.pt after a good training run so others can evaluate it. git rm --cached plus a .gitignore rule stops tracking it, but the blob stays in branch history, so every clone still downloads it; removing that needs a history rewrite and force push, which disrupts everyone on the branch. Catch it before the push.
A pattern that keeps reproducibility without the binary in git: the training job writes the checkpoint to object storage under a run id. A small committed freeze JSON records the run id, the checkpoint's sha256, the threshold and the preprocessing. Evaluation and inference download the weights and refuse to run if the hash differs. Then git holds a few hundred bytes, and the exact model is still pinned.
Add the ignore rule (e.g. checkpoints/*.pt) when you create the checkpoint directory, before anyone has a reason to commit into it.