What is the desired number of positive and negative abstracts from scientific articles for creating such a classification model?