关于sft阶段中数据拼接的问题

您好，论文3.2中有提到将训练数据随机拼接到4k token的长度，请问是指将sft数据拼接成（q0,a0,q1,a1,...）的形式后只计算answer部分的loss吗？
非常感谢大佬们的工作~

The issue asks how section 3.2 of the paper handles randomly concatenated SFT data and answer-only loss, but names no repository file or test. Start by reading section 3.2 and locating the repository's SFT training implementation. Done means documenting a maintainer-confirmed explanation of the data format and loss calculation.

Written by the indexing model from the issue text.