[2310.14633] Extending Input Contexts of Language Models through Training on Segmented Sequences