Cancelled training generated best.pt, but Predict shows no model

Discussion
,
1

Hello,

I trained a model on Ultralytics Platform, but the training was cancelled and the resulting model is not available in the Predict tab.

Experiment Information
• Platform project: fan-phoenix/hardcore-narwhal
• Experiment: exp-5
• Experiment URL:
https://platform.ultralytics.com/fan-phoenix/hardcore-narwhal/exp-5
• Model: YOLO11s
• GPU: NVIDIA H200
• Ultralytics version: 8.4.95
• Python: 3.11.10
• PyTorch: 2.4.1+cu124
• Training progress: 106 epochs
• Training duration: approximately 0.933 hours

Problem
The console shows that both best.pt and last.pt were created:

106 epochs completed in 0.933 hours.
Optimizer stripped from /runs/detect/fan-phoenix/hardcore-narwhal/exp-5/weights/last.pt, 19.2MB
Optimizer stripped from /runs/detect/fan-phoenix/hardcore-narwhal/exp-5/weights/best.pt, 19.2MB

Validating /runs/detect/fan-phoenix/hardcore-narwhal/exp-5/weights/best.pt...
Ultralytics 8.4.95 🚀 Python-3.11.10 torch-2.4.1+cu124 CUDA:0 (NVIDIA H200, 143166MiB)
YOLO11s summary (fused): 101 layers, 9,415,896 parameters, 0 gradients, 21.3 GFLOPs
Platform: Training cancelled from Platform ⚠️

The final metrics were approximately:
• Precision: 0.894
• Recall: 0.816
• mAP50: 0.850
• mAP50-95: 0.688

However, the Predict tab says that no model is available. I cannot find or select best.pt or last.pt for prediction or continued training.

Request
Could you please check:

  1. Whether best.pt was successfully uploaded to Platform;
  2. Whether the checkpoint can be associated with this experiment;
  3. Whether the model can be restored so it appears in the Predict tab;
  4. Whether last.pt can be used to resume training if the model cannot be restored.

According to the Platform callback documentation, cancelled training should still attempt to upload the best checkpoint. If the upload fails, the log should contain:

Model will not be available for download on Platform (upload failed)

Thank you for checking this issue.

2

Hello @fan1

Was this cancellation triggered automatically or did you do it from the Platform UI?

3

Hi @fan1,

We investigated this exact run. The training was stopped automatically when the account balance was exhausted; it was not cancelled manually from the Platform UI.

Your uploaded best.pt was still present in storage and we validated that it is a healthy YOLO detection checkpoint. We have restored its association with fan-phoenix/hardcore-narwhal/exp-5, so it should now be available for download and in the Predict tab.

We also deployed a Platform fix so successfully uploaded checkpoints remain associated with the model when training succeeds, fails, is cancelled, or is interrupted. Periodic checkpoint uploads remain asynchronous and occur at the existing 15-minute cadence; the final upload remains blocking. Platform exposes one canonical best.pt, and each newer valid checkpoint overwrites the older one rather than creating multiple downloadable checkpoints.

Please refresh the experiment page and let us know if you still cannot access the restored checkpoint.