NExtLong: Toward Effective Long-Context Training without Long Documents | Read Paper on Bytez