Learn AI /Training and Model Adaptation /Checkpoint
Checkpoint
- Intermediate
- Training and Model Adaptation
- Technical readers
Also called: model checkpoint, snapshot.
Back to the milestone you came from
In five seconds
What it is
A saved copy of a model’s state at a particular point in training or adaptation, so work can be resumed, compared or released later.
In plain words
Think of it as a saved game for model training. It lets you restart from a particular point or compare different stages instead of beginning again.
An everyday example
At a food distribution point, a team is preparing parcels for 8,000 people. Every hour, the coordinator saves the current packing spreadsheet, the stock counts and notes about which loading rules were being used. If the laptop stops working, the team can reopen the latest saved package and continue instead of starting the day again. Later, they can compare the 10:00 and 14:00 saves to see which plan led to fewer shortages.
A warehouse plan is human-readable and can be checked cell by cell. A model checkpoint is usually a large collection of numerical tensors and training state, not a clear explanation of what the model will output. Also, real training may save many specialised items such as optimiser statistics, which have no simple equivalent in the spreadsheet story.
Remember thisA checkpoint captures model state at a particular training point. Full checkpoints support reliable resumption; weights-only artefacts may support inference but are not equivalent to a resumable training state.
Understand it
Understand it
Why it matters
If you hear that a team has 'the checkpoint', that does not automatically mean they have a finished, approved model. A checkpoint may be an early, unstable or task-specific training state. This matters for procurement, testing, risk review and reproducibility. You may need to ask: Which checkpoint was used? Was it the best one on validation data? Does it include enough information to resume training? Was it merely saved during experimentation, or was it packaged and documented as a release? Clear answers help avoid using the wrong model state, losing work after failures, or comparing systems unfairly.
How a checkpoint is created and used
Show that a checkpoint is a saved training state that can be used to resume training, compare stages or release a selected stage.
01
1. Training updates the model
The model’s weights change step by step as data passes through training.
Draw a box labelled 'Training run' with a progress indicator and small changing number grids to suggest weights updating.
02
2. Save a checkpoint
At a chosen step, the current state is written to storage.
Draw an arrow to a box labelled 'Save at step 12,000' with icons for weights, optimiser state and step number.
03
3. Stored checkpoint
The result is one checkpoint file or a group of files representing that moment.
Draw a storage box or stacked files labelled 'checkpoint-12000'.
04
4. Use it later
The saved state can be loaded to continue training, compare performance, or become the basis of a release.
From the stored checkpoint, draw three outgoing arrows to small boxes labelled 'Resume training', 'Evaluate on validation set' and 'Package as release'.
A left-to-right step diagram with four boxes. It starts with training in progress, then saving a snapshot, then storing one or more checkpoint files, then three possible uses from the saved state: resume, evaluate or publish. Read from left to right.
Hover or tap each step to reveal a one-line note about what is usually saved and what each later use requires.
Describe this without the diagram
Step 1 shows a training run where model weights are changing over time. Step 2 shows a save action at a specific training step, storing the current weights and often optimiser state and step number. Step 3 shows the saved result as a checkpoint file or group of files. Step 4 shows three later uses of that saved state: resuming training after interruption, evaluating that stage against other saved stages, or packaging one selected stage into a released model version. The main message is that a checkpoint is a saved training state, not automatically a public release.
With reduced motion: A static four-step horizontal diagram with labelled arrows: Training run, Save at step 12,000, Stored checkpoint, and three branches for Resume training, Evaluate, and Package as release.
During training, a model’s parameters change many times. A checkpoint is a saved record of that state at one moment. It usually includes the model weights, and often also optimiser state, training step number and other settings needed to continue training.
What goes in is the current training state: the numerical weights and, if the goal is to resume training, the extra bookkeeping from the training process. What happens is simply a save operation to storage. What comes out is a file or set of files that can later be loaded.
People use checkpoints for several different jobs. They can restart after a machine failure, compare earlier and later training stages, pick the best-performing stage on a validation set, or publish a particular stage for others to use. In adaptation work such as fine-tuning, a checkpoint may be saved every few steps or at the end of training.
This term is often confused with model version and model weights. A model version is a named release for users, which may be based on one checkpoint but also includes packaging, testing and documentation. Model weights are the parameter values themselves as an artefact; a checkpoint may include those weights plus extra training state.
How “Saving the plan during warehouse packing” maps onto the real thing
| In the story | In the AI system |
|---|---|
| hourly saved package of spreadsheet, stock counts and notes | checkpoint containing model state and training metadata |
| reopening the latest save after a laptop failure | resuming training from a checkpoint after interruption |
| comparing the 10:00 and 14:00 saves | evaluating different training stages to choose a better checkpoint |
A worked example
An information management officer in a Resident Coordinator's Office is overseeing adaptation of a model that tags comments on a draft results framework before its first reading.
Starting from: The run creates a full checkpoint after each validation pass and a weights-only artefact for candidate inference tests.
- The first full checkpoint records the model parameters, optimiser and scheduler state, random states, tokenizer and configuration references, and the current training step.
- A later checkpoint performs better on the held-out validation set, so it becomes the leading candidate.
- The run is interrupted. Training resumes from the full checkpoint rather than the weights-only artefact, preserving the saved optimisation state.
- After validation, the officer selects the strongest checkpoint and packages its weights separately for controlled evaluation.
Result: The selected weights are used for evaluation, while the corresponding full checkpoint is retained so the adaptation run can be resumed or examined later. The checkpoint itself is not treated as the cleared model release.
In a United Nations workplace
A training officer in an evaluation office is overseeing adaptation of a multilingual model to route monitoring reports after translations have accumulated in the queue. The front office wants a tested configuration before close of business.
Checkpoints allow the officer to compare intermediate results and resume an interrupted run. A full checkpoint preserves the wider training state, while a weights-only save may be sufficient for evaluation but not for continuing adaptation with the same optimisation behaviour.
Before you try thisMonitoring reports may contain personal or confidential material. Do not upload source documents, training logs or checkpoints to any external tool or storage location not approved for that material. Apply the required access and retention controls because checkpoints can preserve learned information from the training data.
Where you may hear it
- A vendor demo to an ICT governance board
- A technical review before a model release is cleared
- A handover call after an interrupted fine-tuning run
- A procurement meeting on storage and recovery requirements
Avoid confusion
Avoid confusion
Often confused with
Model version
A particular identified state of a model or model-backed system, with documented changes, capabilities and checks for use.
Model weights (as a released artefact)
A released file of learned numerical settings that lets compatible software run a trained model without repeating the original training process.
Knowledge cutoff
The approximate point up to which a model’s training-derived background is expected to be current, while newer events and facts may be absent unless added another way.
Common misconceptions
Common misconception: A checkpoint and a model version are the same thing.
Correction: A checkpoint is a saved training state at a moment in time. A model version is usually a managed release label for users, often with testing, documentation, serving configuration and governance decisions. One version may come from one checkpoint, but the terms are not interchangeable.
Common misconception: A checkpoint only contains the model weights.
Correction: Sometimes a lightweight save contains just weights, but a training checkpoint often also stores optimiser state, step counts, scheduler settings and other metadata needed to continue training properly. Without those extras, you may load the model for inference but not resume training in the same way.
Common misconception: The latest checkpoint is always the best one to deploy.
Correction: Later training is not always better. Performance on validation data can level off or worsen, and other risks can appear. Teams often select an earlier checkpoint because it performed better on agreed evaluation criteria.
Quick check
Which description best matches a checkpoint in model training?
A checkpoint is a saved training or adaptation state. It can resume work, compare stages or support a release. Other options are a release name, an inaccurate cutoff claim, and runtime prompting.
Go deeper
Go deeper
The simple version
Think of it as a saved game for model training. It lets you restart from a particular point or compare different stages instead of beginning again.
The technical version
A checkpoint is a serialised snapshot of training state at a specific step or epoch. Depending on the framework and purpose, it may include model parameters, optimiser moments, learning-rate scheduler state, random number generator states, gradient-scaler state, tokenizer or config references, and the global step. For inference-only use, some artefacts save only weights. For resumable training, the fuller state matters because restarting from weights alone can change optimisation behaviour. In fine-tuning and parameter-efficient methods such as LoRA, the checkpoint may store only adapted parameters or adapters rather than the whole base model, depending on how the run was configured.
Reviewed external resources
EncyclopaediaCheckpoint (opens in a new tab)Wikipedia
This link gives a broad introductory overview of checkpointing and suits readers who want a quick orientation before going deeper.
Intermediatelink checked 2026-07-31
CourseLLM Course (opens in a new tab)Hugging Face
This course link is useful for learners who want practical training context and suits readers comfortable with hands-on machine learning workflows.
Deeperlink checked 2026-07-31
CourseShort Courses (opens in a new tab)DeepLearning.AI
This course link offers short, applied lessons and suits readers who prefer guided examples over reference-style explanation.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Checkpoint (opens in a new tab)YouTube search
This is a YouTube search results page, so it helps readers browse multiple video explanations and suits people who learn well from varied visual teaching styles.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Checkpoint (opens in a new tab)Reddit search
This is a Reddit search results page, so it helps readers see community questions and practical confusion points and suits readers who want informal discussion alongside formal sources.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
A checkpoint captures model state at a particular training point. Full checkpoints support reliable resumption; weights-only artefacts may support inference but are not equivalent to a resumable training state.