Getting started

What a release contains

• Video: one MP4 per sensor stream per episode, constant frame rate, timestamped against a common episode clock.
• Structure: episode records with task, environment, duration, contributor pseudonym and outcome.
• Annotations: frame-accurate subtask boundaries, tool and object metadata, outcome labels.
• Provenance: capture window, environment class, protocol version, processing version and schema version per episode.
• Licence: the exact terms the release is delivered under, as a file inside the release.
• Documentation: the schema version, the changelog and the known limitations of that specific version.

Get access

Access is granted per dataset, version and licence scope. Once the licence is signed you receive an API key and a download endpoint scoped to the versions you licensed.

curl -H "Authorization: Bearer $CURATRIX_API_KEY" \
  https://api.curatrix.de/v1/datasets/electrical-assembly/versions/1.3.0/manifest \
  -o manifest.json

Verify what you downloaded

Every release ships a checksum file. Verify before training: a partial download is far more expensive to discover after a training run than before one.

sha256sum -c checksums.txt

Load the first episode

from curatrix import Dataset

ds = Dataset.open("electrical-assembly", version="1.3.0")
episode = next(ds.episodes(outcome="success"))

print(episode.id, episode.task, episode.duration_s)
for step in episode.subtasks:
    print(f"{step.start_s:7.2f} -> {step.end_s:7.2f}  {step.label}")