Skip to main content
The Run tab trains a model on a use case, or scores one, without leaving the dashboard. You pick a model, the platform checks it, you configure the experiment and click Start. It is the quickest way to a first experiment. For full control in code, use the notebook instead: both start the same kind of experiment.

​
Open the Run tab

Join the use case first: see Join. Then open the use case and do one of these:
  • Click Run in the use case’s tab bar.
  • Click Start training in the Overview’s On this page sidebar.
  • Click Train a model next to Model Leaderboard on the Leaderboard tab.
  • On an empty Training tab, click How to train a model.
Each opens the Run your model panel over the page. Closing it with Esc, the close icon or a click outside keeps what you chose, so you can look at the use case and come back. The panel keeps one state per use case. The top of the panel shows the use case, its dataset (for example 134 images · 3 labels) and its secure environments, with how many are online. Hover over the secure environments to see what each experiment gets there: GPU, CPU cores and memory. The panel runs a notebook that tracebloc starts for you on first use. If it reports that no notebook can be started, close a notebook you have open on another use case, then try again.

​
Step 1: Select your model

  1. Open Model and pick a template from the model zoo, or choose Upload a .py from your computer… (a .py or .zip file). You can also drop the file onto the field. The list shows only the templates that fit the use case’s task; View code opens the chosen template on GitHub.
  2. Optional: open Pretrained weights and pick the zoo’s weights, or upload a .safetensors, .pth or .pkl file. Without weights, training starts from scratch. A template from the zoo runs with the zoo’s weights or none: a weights file chosen next to a template is not loaded.
The checks start as soon as you choose a model. There is no button to press: Model checks passed means you can go on. When a check fails, a table lists each finding with the parameter, what was expected, what was found and the fix. Some fixes have a button that applies them for you, for example writing the right output_classes into the model file. Fix the file and choose it again: changing a file runs the checks again. Raw output shows the checker’s own text. If Trainable keeps failing for no reason in your model, run it again; if it still fails, the hint names the cause, for example a notebook package older than the check.

​
Step 2: Configure your run

Step 2 opens once the model passes its checks. Choose Train or Inference at its top right:
  • Train trains the model across the secure environments. With pretrained weights, it fine-tunes them. The settings are a short Python cell: training is your model, linked to the dataset, and typing training. lists every setting. They are the same settings as in the notebook: see Hyperparameters. Click Validate configuration. The result line shows the plan that applies, for example the cycles, epochs and batch size, and its estimated cost in PetaFLOPs (PF). After an edit it reads Edited, not applied · validate again.
  • Inference scores the model on the use case’s test data without training it. There is nothing to configure: the line shows the estimated cost. Without weights, it scores the zoo’s own weights, if any, or random weights as a baseline. Inference counts against your team’s daily submissions, which reset at 00:00 UTC.

​
Step 3: Start

Click Start. When Start is greyed out, hover over it to see why, for example Validate the configuration or No edge holding this data is online. If some secure environments are offline, you can still start: a warning names the data that is left out. Once the experiment starts, the panel shows its key:
  • After Train, click Monitor in training to follow it on the Training tab. See Start Training for pausing, resuming and stopping it.
  • After Inference, click See the scores. The result is a submission labelled Inference only on the Leaderboard.
If the platform refuses the start, the panel shows why. Not enough petaflops left for this run means your team’s compute budget is used up: see When the compute budget runs out.

​
Run again

After a start, Run again replaces Start. It takes you back to step 2 with the same model, its checks, its weights and your settings, so you can change a setting and start another experiment. After Train, click Validate configuration again before Start. The first experiment keeps running. Start over (on the panel’s title line) clears the model, the weights and the settings. It does not stop an experiment you started: stop that from the Training tab.

​
When to use the notebook instead

The Run tab covers one model file, its weights and the training settings. Use the notebook, locally or in Google Colab, when you want to script several experiments, keep your code in your own files, or use SDK calls the panel doesn’t offer.

​
Need Help?