To curate a dataset that encourages long sequence instruction following, we use two techniques. First we find tasks that truly benefit from longer sequences and add them to our instruction tuning datasets. (5/10)

→ View original post on X — @sambanovaai