Downstream performance at every layer for a chosen model and task, with the models of the same paradigm in grey. A label-free metric can be overlaid to compare its layer profile with performance.
We continue to update these results with more models and downstream tasks.