> ## Documentation Index
> Fetch the complete documentation index at: https://docs.dreamscalelabs.com/llms.txt
> Use this file to discover all available pages before exploring further.

> ## Agent Instructions
> Treat the Dreamscale CLI and Python SDK as the supported public integration surfaces.
> Prefer `dreamscale run <robot>` for supported robots and `dreamscale.connect(model)` with `predict` or `next_action` in Python.
> Do not infer or recommend internal control-plane APIs.

# Training

> Fine-tune π0.5 on your robot's LeRobot dataset, then run it as a deployment.

Training fine-tunes π0.5 on your robot's demos, on our GPUs. One command trains it,
and one more turns it into a [deployment](https://docs.dreamscalelabs.com/deployments.md) you can run.

## What you need

* The latest CLI: `uv tool upgrade dreamscale`.
* Credits on your account: you see the price and approve it before anything starts.
* A LeRobot dataset (v2.x or v3.0), in a local folder or on Hugging Face.

## 1. Train

```bash theme={null}
dreamscale train ./towel-fold
```

From Hugging Face: `dreamscale train hf://your-org/towel-fold`.

The CLI checks the dataset, drops faulty episodes, and shows the price and time.
Say yes and it uploads, starts training and shows progress on one line. It ends
with **Ready**. Ctrl-C stops watching, not training: `dreamscale train show tr_...`
picks it back up.

The defaults choose the base model and training length for your data. To set a
budget instead, add `--max-usd 50`. For a private repo, run
`dreamscale integrations hugging-face setup` once.

## 2. Deploy it

```bash theme={null}
dreamscale train deploy tr_...
```

The build takes a few minutes and prints the command that runs it on your robot.
To rename it or keep it warm, see [Deployments](https://docs.dreamscalelabs.com/deployments.md).

## Manage runs

```bash theme={null}
dreamscale train list
dreamscale train cancel tr_...   # you pay only for the steps it finished
dreamscale train rm tr_...       # permanent; a running one is canceled first
```

The dashboard's Training page shows the same runs.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.