Submissions are evaluated using the **pairwise accuracy**, which evaluates how well are the models able to align with human preferences, which text should be preferred in a given pair. 

In summary, **the proposed solutions should focus on** minimizing the number of wrongly chosen texts from each of the pairs.
The metric does **NOT** care about computation timing, however please have in mind the Kaggle execution time limits: 12 hours for CPU and GPU notebook sessions and 9 hours for TPU notebook sessions.