Corrupted prediction after model export

Hi. I’m trying to fine tune “bart-large-cnn” to find the pros and cons of a product described in the text (following this collab Google Colab).

The problem is that after exporting fine-tuned model the predictions of the same text (that was used prior to exporting) are bad and include corrupted symbols (see screenshot). Could anyone point out what I’m doing wrong? I’m new to and deep learning.