Examples
Worked examples
- Is an instance
A checkpoint file named 'model_step_120000.safetensors' with corresponding optimiser state.
- Is an instance
A research release of 36 intermediate checkpoints from a single 70B training run for studying emergence.
Counter-examples
Looks similar, but isn't
- Not an instance
A model card without any saved weights.
- Not an instance
A configuration file with no learned parameters.
Editorial commentary
A model checkpoint is a saved snapshot of a model’s parameters — and often its optimiser state — at a specific point during training, identified by a training-step number or an internal tag and serialised to a file format such as safetensors or PyTorch’s .pt. Checkpoints exist primarily as internal training infrastructure: they let a run be resumed after interruption, rolled back if training diverges, or compared across steps to study how a capability emerges over the course of training.
How this differs from a version or a lineage record
Not every checkpoint becomes a public release. A training run may produce thousands of checkpoints, of which only a handful are ever externally versioned (see model versioning) and released. A checkpoint is the artefact; a version is the externally-facing label some checkpoints receive; and lineage is the graph describing how one checkpoint, or released model, relates to another. A fine-tune-lineage record typically names the specific checkpoint a fine-tuning run started from, and the checkpoint it produced.
Why it matters for reproducibility
For research that fine-tunes or otherwise builds on an existing model, citing the exact checkpoint identifier — not just the model family name — is the equivalent of citing a specific dataset version: without it, a downstream researcher cannot be certain they are building on the same starting point.
References
- See also: Model fine-tune lineage, AI and ML research outputs.
Also known as
weights checkpoint · training checkpoint
Machine-readable encodings
Use in your systems
<role vocab="credit"
vocab-identifier="https://casrai.org/dictionary/"
vocab-term="Model checkpoint"
vocab-term-identifier="https://casrai.org/dictionary/term/model-checkpoint" />{
"@context": "https://schema.org",
"@type": "DefinedTerm",
"@id": "https://casrai.org/dictionary/term/model-checkpoint",
"name": "Model checkpoint",
"identifier": "https://casrai.org/dictionary/term/model-checkpoint",
"description": "A saved snapshot of a model's parameters (and optionally optimiser state) at a specific point in training, identified by a step number or version tag and serialised to a file format such as safetensors or .pt.",
"inDefinedTermSet": "https://casrai.org/dictionary/domain/ai-ml-research-outputs#set",
"url": "https://casrai.org/dictionary/term/model-checkpoint",
"sameAs": [
"weights checkpoint",
"training checkpoint"
],
"license": "https://creativecommons.org/licenses/by/4.0/",
"publisher": {
"@id": "https://casrai.org/#organization"
},
"author": {
"@id": "https://casrai.org/#editorial-team"
},
"datePublished": "2026-05-21T02:22:51",
"dateModified": "2026-08-22T15:43:52",
"inLanguage": "en-GB",
"isAccessibleForFree": true
}
