After #13, we can add one more encoder model, e.g., EuroBERT. It should be relatively easier than autoregressive models.
