Creating Training Jobs
Where: Model Training → Task Management in the left navigation, then click Create Training Job.
Prerequisites
| Requirement | Details |
|---|---|
| Role | FAE or R&D engineer |
| Dataset version | At least one dataset version with status Exported |
| Compute quota | An available training compute quota |
Preparing Training Inputs
A training job takes dataset versions as input, not data assets or curation projects.
- Only Exported versions can be selected; versions that are exporting or failed to export are not listed.
- You can train on several versions at once, but they must all be in the same format (all HDF5 or all LeRobot).
- To combine several batches of data, you can also merge them into a new dataset version first and train on that single version.
Input quality sets the ceiling
Before training, make sure failed demonstrations have been excluded and every item has been reviewed. You can export a small version first, train on it, and export an early checkpoint to check direction.
Configuration Options
| Option | Required | Description |
|---|---|---|
| Job name | ✓ | Shown in the job list; any characters allowed |
| English name | ✓ | Letters, digits, and underscores only; used to name artifacts |
| Training data | ✓ | One or more exported dataset versions in the same format |
| Model family | ✓ | ACT, EasyDP, Pi 0.5, R0, and others; determines the network architecture and default hyperparameters |
| Compute tier | ✓ | Light / Standard / High-performance |
| Training strategy | — | Fixed by the platform as "long-horizon, quality-first"; shown but not editable |
| Notes | Record the purpose of this run and what changed in the data, to make comparisons easier |
Compute Tiers
| Tier | Reference configuration | Best for |
|---|---|---|
| Light | 2 × GPU (80 GB class) · 16 vCPU · 128 GiB memory | Small datasets and first experiments |
| Standard (recommended) | 4 × GPU (80 GB class) · 32 vCPU · 256 GiB memory | Typical training jobs |
| High-performance | 8 × GPU (80 GB class) · 64 vCPU · 512 GiB memory | Large datasets and larger models |
The tier affects training speed and batch size, not the step budget or strategy.
Submission and the Parameter Snapshot
When you click Submit, the platform locks the following parameters:
| Parameter | Source |
|---|---|
| Training steps | Unified strategy (currently 80,000 steps) |
| Learning rate and schedule | Model family template |
| Batch size | Model family template and compute tier |
| Optimizer | Model family template |
| Random seed | Generated at submission |
| Dataset versions | Selected at submission |
Later changes to a training template don't affect jobs that have already been submitted, so past training runs remain reproducible.
Queueing
When no compute is available, the job enters Queued and the page shows its position in the queue; it is never shown as started before it actually is. Training begins automatically once compute becomes available. Don't resubmit the same job to try to speed it up.
Training Duration
Training time depends on GPU model, batch size, data loading, validation frequency, and queueing, so the platform doesn't promise a fixed duration. The job details page shows the current step and elapsed time, which you can use to estimate the time remaining.
Common Issues
| Issue | Resolution |
|---|---|
| The training data list is empty | No dataset versions have been exported yet; complete an export first |
| Selecting several versions shows a format mismatch | Only versions in the same format can be combined |
| You want to change the number of training steps | The step budget is set by the unified strategy. To check direction early, export a checkpoint. |
| You want to change the configuration after submitting | The parameter snapshot is locked. Stop the job and create a new one. |

