---

name: data-preparation description: "Prepare, validate, register, and troubleshoot Dexbotic DexData datasets and safe dataset conversions." metadata: disco-role: operating disable-model-invocation: true license: MIT

Dexbotic data preparation

Use this route when the task is about converting demonstrations, designing a DexData JSONL schema, registering a dataset, computing action/state normalization metadata, or diagnosing data-loader failures. Keep model/trainer selection in training, HTTP policy serving in inference-serving, and physical robot topology in evaluation-deployment.

Operating sequence

1. Decide the action/state convention before conversion: dimensionality, delta versus absolute dimensions, periodic dimensions, gripper convention, and camera order.
2. Produce one JSON object per frame in an episode JSONL file. Keep image/video references and frame indices aligned with the same frame.
3. Validate paths, finite numeric values, frame indices, prompt presence, and consistent state/action dimensions with scripts/validate_dexdata.py before registering the dataset.
4. Register a dataset dictionary with register_dataset(...), including annotations, optional data_path_prefix, positive frequency, and action metadata. Import the registration module before constructing DexDataset.
5. Generate or load norm_stats.json alongside the checkpoint/data contract; never silently reuse statistics from another action space.
6. Run a small loader inspection before expensive training. A missing image, malformed JSONL line, or bad action horizon is a data issue, not a model issue.

Read DexData and registration, conversion boundaries, and troubleshooting for details. The bundled validator is intentionally local, deterministic, and read-only unless an explicit report path is supplied.