How can i use LayoutXLM Pretrained for sequences with length > 514 ?

[Alirez1043]

I trained a LayoutXLM model with max_sequence_len = 512. When using the model on a document with a higher number of detected boxes, it predicted the last boxes as the 'other' class due to its limitation in handling a larger number of boxes.

Additionally, I experimented with a model trained with max_sequence_len = 714 and max_position_embedding = 714, but the results were unsatisfactory. During training with a length of 512, the accuracy reached 93%, whereas with 714, the accuracy dropped to 72%.

I am seeking guidance on how to address this issue. What adjustments can I make to improve the model's performance, especially when dealing with a larger number of detected boxes?