Rules
General Rules
The general rules applied to all tracks.
The rules here apply to all the submissions, including
Track Rules
The track-specific rules for each of the tracks.
Constrained
A fair comparison between frozen representations by enforcing the same downstream model in each task. Only a few hyper-paramters for training are allowed to tuned.
Less-constrained
A comparison between frozen representations with customized but limited-resource downstream models. The details of downstream models are reported along with submissions.
Unconstrained
The track does not limit any approach for solving SUPERB tasks as long as it in principle utilizes only a single shared model with limited task-specific parameters.
Submission example
Download an example of submission.
The prediction output of each task should be put in seperated folder sepified by the task. Finally, please put all the task folders in one folder called predict and submit a zip archive. You can download an example for submission from this Link.