need ignore_empty_think when training Qwen3.5？

Checklist / 检查清单

• I have searched existing issues, and this is a new question or discussion topic. / 我已经搜索过现有的 issues，确认这是一个新的问题与讨论。

Question Description / 问题描述

当训练千问 3.5 纯文本的部分，例如同时使用Qwen3的推理和非推理语料，是否需要加上这个标签来跟原始的基座能力对齐？
--loss_scale ignore_empty_think \

When training the Qwen 3.5 base model on pure text data—where both reasoning corpus and non-reasoning corpus coexist—do we need to add this label to align with the original base model capabilities？

Start with the Qwen 3.5 training setup and the --loss_scale ignore_empty_think option referenced in the issue. Compare mixed reasoning and non-reasoning pure-text data with the original base-model behavior; done means a project-confirmed answer about whether the option is required.