Finetune on longer texts

I'm looking to find an effective way to fine-tune this for a classification task with longer text rather than sentences.
The dataset structure just like those on NLI datasets with
**text 1** will be longer such as a document with many paragraphs
**text 2** will be shorter such a paragraph describing a category.
**score** will be a score from 1 to 5

(1) In this situation, i'm not sure if the effects of the length of text on the performance of the classification performance as well as the quality of the embedding generated for this domain.
(2) what would be a optimal size of text length in this method, or does it align with the token size (1024) from the original model (BERT)
(3) In case of using the existing SBERT pre-train model, what would be the optimal size of text? or would it be useful for multiple sentences?

The issue does not identify any files, tests, or entry points. Start by locating the fine-tuning and text-length handling paths, then define an approach for longer document and paragraph inputs and validate its impact on classification and embeddings.

Written by the indexing model from the issue text.