A saved snapshot of a model's weights at a point during training, enabling recovery, comparison across training stages, and selection of the best-performing version.
A checkpoint is a saved copy of a model's weights at a specific point during training. Rather than waiting until training finishes to save anything, most training pipelines save checkpoints at regular intervals — every N steps or every epoch — so that progress is never fully lost to a crash, and so that multiple versions of the model can be compared.
Training long-running models without checkpoints is a bet that nothing will go wrong. Checkpointing turns a catastrophic failure into a minor setback. When LinkjeBERT trained for 10 epochs over 7 days, checkpoints meant each epoch's weights were preserved. If a hardware fault had hit on day 6, training could resume from the last saved checkpoint rather than from scratch.
Checkpoints also enable early stopping: you save a checkpoint at each validation step, and after training finishes (or stops) you load whichever checkpoint had the best validation score, not necessarily the final one. The final epoch's weights are often slightly overfit; an earlier checkpoint may generalise better.
Selecting the right checkpoint is part of the training workflow. DEJAN's model training pipelines monitor validation metrics — F1, precision, recall — at regular intervals and flag the best-performing checkpoint automatically. The model shipped to production is that checkpoint, not necessarily the one from the last training step.
When DEJAN publishes models to Hugging Face (dejanseo/LinkBERT, dejanseo/Intent-XS, etc.), what is published is a checkpoint: the saved weights, configuration, and tokeniser from the best training run. Anyone who downloads it gets a frozen snapshot of the model at its best-validated state.
