Hi there,
If I understood your situation, you have a context, questions and answers, but your context is too big to feed the model in one passage.
I suggest you to take a look on Hugging Face’s question answering example notebook. They manage to solve this problem splitting up the context in several parts, when necessary. During training, if the context split does not contain the answer, they point the answer to the CLS token. Finally, during evaluation, they search for the answer in all splits and consider the most likely answer.