通过分析模型词汇中的异常中文标记，推测其训练数据是否被不良信息污染。
Speculating LLMs' Chinese Training Data Pollution from Their Tokens
- 基于GPT词表定义并分类有害中文标记（PoC tokens）。
- 用微调模型检测词表中含色情或赌博内容的标记，发现超23%长中文标记有问题。
- 可反推训练数据中特定网页占比，适合关注AI安全与数据溯源的研究者。
LLM训练数据的基本单元是标记（tokens）。研究表明，GPT（4o/4o-mini/o1/o3/4.5/4.1/o4-mini）词表中存在大量代表中文短语的标记，其内容涉及色情或网络赌博。本文旨在定位这些污染型中文标记（PoC tokens），并探究其与训练数据的关系。（1）基于GPT词表，提出PoC token的形式化定义与分类体系。（2）通过微调大模型，结合语义与搜索引擎信息，构建PoC token检测器。（3）基于标记出现情况（token ID）推测训练数据污染状况。在GPT及其他23个LLM上的实验表明，此类标记广泛存在，其中GPT词表最严重：超过23%的长中文标记（含两个以上汉字）属于色情或网络赌博类别。通过对C4、Pile等知名预训练数据集的验证，确认推测方法的有效性。以GPT-4o为例，推测其训练数据中与“Yui Hatano”相关网页占比约为0.5%。
原文摘要 · Abstract (English)
Tokens are basic elements in the datasets for LLM training. It is well-known that many tokens representing Chinese phrases in the vocabulary of GPT (4o/4o-mini/o1/o3/4.5/4.1/o4-mini) are indicating contents like pornography or online gambling. Based on this observation, our goal is to locate Polluted Chinese (PoC) tokens in LLMs and study the relationship between PoC tokens' existence and training data. (1) We give a formal definition and taxonomy of PoC tokens based on the GPT's vocabulary. (2) We build a PoC token detector via fine-tuning an LLM to label PoC tokens in vocabularies by considering each token's both semantics and related contents from the search engines. (3) We study the speculation on the training data pollution via PoC tokens' appearances (token ID). Experiments on GPT and other 23 LLMs indicate that tokens widely exist while GPT's vocabulary behaves the worst: more than 23% long Chinese tokens (i.e., a token with more than two Chinese characters) are either porn or online gambling. We validate the accuracy of our speculation method on famous pre-training datasets like C4 and Pile. Then, considering GPT-4o, we speculate that the ratio of "Yui Hatano" related webpages in GPT-4o's training data is around 0.5%.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。