This dataset is built from scrapped speeches on the Bank of International Settlements thanks to this repo :
https://github.com/HanssonMagnus/scrape_bis. The dataset is made of 12k speeches from 1997 to 2023.
Each pair is built with extracted sentences from speeches, if B is following A then the 'next_sentence_label' is 1 else it is 0.
Negative pairs are built by choosing a sentence from another speech randomly.
