模型基因组：通过指纹识别LLM是从零训练还是衍生
原标题：Model Genome: Fingerprinting Whether an LLM Was Trained From Scratch or Derived
AI 摘要
Hugging Face 博客发布了一篇技术文章，介绍如何通过分析公开的模型配置文件（config.json）、分词器（tokenizer）和权重（embedding CKA）来客观判断一个大型语言模型是从零训练还是基于开源基础模型衍生而来。文章指出，架构指纹和分词器重叠是主要证据，而权重相似性因旋转不变性等原因不可靠。作者将该方法应用于韩国九个组织的公开基础模型，并提供了在线工具 Model Genome Korea。
正文节选
Model Genome Korea DNA lineage test of Korean LLM & VLM foundation models config.json), tokenizer (vocabulary overlap), and weights (embedding CKA) — and combined them into a single at-a-glance genotype. Along the way we hit two instructive traps: row-wise embedding cosine is useless because of rotational invariance, and even CKA cannot cleanly separate continued-pretraining from from-scratch — so config + tokenizer remain the primary evidence. We applied the exact same yardstick to the public f