Models performances right out of the box


I had a question about model performance on glue- it seems from the website that bert-base was fine-tuned on the task before these numbers were reported.

Here are the results I got using untrained bert-base models:

Is there a way to check which models have already been fine-tuned on a task and get their weights?