https://cybersecurityerial.github.io/echo_blog/tags/initialization/Recent content in Initialization on 我的技术博客Hugozh-cnWed, 06 May 2026 13:09:10 +0800https://cybersecurityerial.github.io/echo_blog/posts/llm-theory-02-initial-setup/Wed, 06 May 2026 13:09:10 +0800https://cybersecurityerial.github.io/echo_blog/posts/llm-theory-02-initial-setup/<blockquote> <p>本文从训练模型要考虑的第一性原理（稳定性和速度）出发，探讨了 LLM 预训练中的初始化设定问题。</p> </blockquote> <h2 id="参考文献">参考文献</h2> <ol> <li><a href="https://spaces.ac.cn/archives/11340">MuP之上：1. 好模型的三个特征</a></li> <li><a href="https://spaces.ac.cn/archives/11605">https://spaces.ac.cn/archives/11605</a></li> <li><a href="https://spaces.ac.cn/archives/11647">https://spaces.ac.cn/archives/11647</a></li> <li><a href="https://spaces.ac.cn/archives/11729">https://spaces.ac.cn/archives/11729</a></li> </ol>