What Is Model Selection in Training? Choosing the Best Checkpoint
A training run produces many saved checkpoints along the way — model selection is the process of choosing which one actually deserves to ship.
Choosing among many saved snapshots
Model selection is the process of choosing which saved checkpoint from a training run — out of potentially dozens saved at different points — actually represents the best version of the model to use going forward, rather than automatically defaulting to whichever checkpoint happens to be the very last one saved.
Why the last checkpoint isn't always the best one
Because of overfitting risk, a model's performance on genuinely new data can peak at some point during training and then gradually decline afterward, even while training loss keeps improving — meaning the checkpoint saved at the very end of a run isn't guaranteed to be the strongest-generalizing one.
What selection is actually based on
Model selection typically compares candidate checkpoints against a held-out validation set and relevant evaluation benchmarks, picking whichever checkpoint scores best on the metrics that actually matter for the model's intended use, rather than picking based on training progress alone.
Frequently Asked Questions
What is model selection in training?
The process of choosing which saved checkpoint from a training run actually represents the best model to use, rather than automatically defaulting to whichever checkpoint was saved last.
Why isn't the final checkpoint automatically the best choice?
Performance on new data can peak at some point during training and decline afterward due to overfitting risk, even as training loss keeps improving — so the very last checkpoint isn't guaranteed to generalize best.