TokLabel (Tokamak Label Suite) is an open-source fusion data labeling platform co-developed by Startorus Fusion and Tsinghua University, built on the open-source Label Studio and released under the Apache 2.0 license. Fusion AI researchers spend roughly 70% of their time on data preparation, and TokLabel targets exactly this stage.It turns raw discharge data into labeled datasets ready for fusion AI. Researchers simply input a shot number to automatically retrieve data and generate labeling tasks — labeling feature moments such as breakdown, disruption, and end on one-dimensional time series, and configuration information such as the last closed flux surface on two-dimensional visible-light images. Labeling results are stored by shot number and directly consumed by downstream model training such as disruption prediction, discharge feature recognition, and configuration recognition, saving roughly 70% of data-preparation time.
| Parameter | Description |
| Labeling types | 1D data (time series), 2D data (images) |
| Automatic data import | Connects to MDSplus / PostgreSQL; input a shot number to automatically retrieve data and generate labeling tasks |
| AI pre-labeling | The model predicts breakdown, disruption, and end feature moments; humans verify and correct, and corrections flow back into training |
| Access control | Supports multi-user collaboration, filling the gap in the open-source edition of Label Studio |
| Storage & retrieval | PostgreSQL + Redis; exported annotations are stored and retrieved/reused by shot number |
Disruption prediction：Train a model on disruption-labeled data to determine whether a shot is disruptive
Discharge feature recognition：Use breakdown / disruption / end time labels to train a model to automatically recognize key discharge events
Configuration recognition：Use last-closed-flux-surface labels on visible-light images to train a configuration recognition model
