Comment by polyterative

I have solved this problem by forcing the model to compress every time by specifying token budgets for the documents