Stop or Cancel a Training Job

Stop training early and save the model, or cancel training and discard its weights.

You can end a training job before it finishes. Stop training early to save the current model weights. Cancel training to discard the weights.

	Stop Training Early	Cancel Training
Use when	The model has converged or already meets your goals.	You do not want a model from the training job.
Model weights	Saved	Discarded
Credits	You pay for the credits used.	You may receive a refund if you cancel early in the job. A refund is not issued after many epochs.

After either action, you can start another training job with the same dataset version.

You can also set a credit cap when you start a job, so it stops on its own once it reaches a credit limit. A cap stop keeps the model weights.

Stopping is cooperative: the job saves its weights and exits at the end of its current epoch. If a job gets stuck and never responds to the stop request, Roboflow ends it after about 8 hours. That job shows as failed instead of stopped, saves no model weights, and its credits are refunded. The same happens when Roboflow stops a job because the workspace ran out of credits and the job never responds to the stop request.

Web App

Stop Training Early

Use early stopping when the training graphs show that the model has converged or already meets your goals, but the job has many epochs left.

Click "Stop Training Early":

Roboflow starts preparing the current model weights. The training job shows that it stopped early:

It may take several minutes for the model to become available. A green checkmark appears next to the model name when it is ready:

For Neural Architecture Search (NAS), training may stop automatically when the model converges. The training results page shows a message when Roboflow skips the remaining epochs.

Cancel Training

Click "Cancel Training" to stop the job and discard its model weights:

HTTP API

The calling API key must have the trainingJob:create scope.

Stop Training Early

Request an early stop for a running training job:

POST https://api.roboflow.com/:workspace/:project/:version/train/stop

curl "https://api.roboflow.com/my-workspace/my-detector/3/train/stop?api_key=$ROBOFLOW_API_KEY" \
  -X POST \
  -H "Content-Type: application/json" \
  -d '{}'

The endpoint is idempotent. It returns success if the job has already stopped.

{
  "success": true
}

Cancel Training

Cancel a queued or running training job:

POST https://api.roboflow.com/:workspace/:project/:version/train/cancel

curl "https://api.roboflow.com/my-workspace/my-detector/3/train/cancel?api_key=$ROBOFLOW_API_KEY" \
  -X POST \
  -H "Content-Type: application/json" \
  -d '{}'

{
  "success": true
}

Cancel is not idempotent. If the job has finished or has already been canceled, the endpoint returns 409 Conflict:

{
  "error": {
    "type": "Conflict",
    "message": "Cannot cancel non-running train job.",
    "code": "CANNOT_CANCEL"
  }
}